TerraMosaic Daily Digest: Mar 2, 2026
Daily Summary
This March 2, 2026 digest synthesizes 399 selected papers from 1,888 analyzed studies. Across this corpus, the strongest contributions couple physically interpretable forecasting with observation-rich diagnostics: model-adequacy-based seismic hazard assessment, reproducible real-time eruption forecasting, spatiotemporal landslide susceptibility mapping with heterogeneous soil fields, and tensor-based UAV/LiDAR measurements that resolve deformation localization in active slopes.
A second front advances risk quantification for built and natural systems by explicitly linking hazard intensity to exposure and design thresholds, including GLOF scenario mapping in Bhutan, climate-conditioned urban flood projections, probabilistic mine-pillar instability analysis, and fault-displacement-informed setback optimization. Together, these studies move from event description toward transferable, uncertainty-explicit geohazard intelligence for planning, mitigation, and rapid response.
Key Trends
- Observation-mechanism coupling now anchors the literature: Concepts & Mechanisms and Detection & Monitoring together account for over three-quarters of selected papers, indicating a field-wide move toward extracting process-level insight from dense multi-sensor observations rather than treating monitoring as a stand-alone endpoint.
- Predictive studies are becoming uncertainty-native: in seismic, volcanic, slope, and subsurface applications, hazard and risk models increasingly report epistemic uncertainty, model adequacy, and scenario envelopes as core outputs, not supplementary diagnostics.
- Landslide research is advancing from static susceptibility to dynamic failure physics: contributions emphasize trigger-lag structure, strain localization, and spatiotemporal heterogeneity, with UAV/LiDAR and physics-based workflows resolving how and where instability evolves through time.
- Multi-hazard analysis is maturing into exposure-aware frameworks: cryosphere, flood, and earthquake-ground-failure studies increasingly couple hazard intensity with downstream exposure and infrastructure vulnerability, improving the utility of results for prioritization and risk governance.
- Methodological innovation is converging on operational robustness: cross-domain AI and geospatial methods are increasingly evaluated under physical constraints, uncertainty quantification, and real-scene validation, reflecting a transition from proof-of-concept novelty to deployable geohazard intelligence.
Selected Papers
This digest features 399 selected papers from 1888 papers analyzed across multiple journals. Each paper has been evaluated for its relevance to landslide and broader geohazard research and includes links to the original publications.
1. Physics-Based Seismic Hazard and Risk Assessment: A New Paradigm for Earthquake Forecasting
Core Problem: Epistemic uncertainty in PSHA is inadequately addressed by logic-tree frameworks, which assume current models sufficiently represent underlying physics, leading to potential systematic hazard bias and underestimated forecast uncertainty.
Key Innovation: SHARP (Seismic Hazard Assessment and Risks with Physics), a new framework that quantifies the collective distance of models from physical and observational constraints using a Model Adequacy Distance (MAD) metric, combining a moment-weighted scoring function and compatibility measures from geodetic observations and seismicity statistics.
2. Axial Seamount Eruption Forecasting Experiment
Core Problem: The need for a transparent, reproducible, and scientifically rigorous framework to test and evaluate the predictability of volcanic eruptions, transforming prediction into a cumulative and testable science.
Key Innovation: Establishes the Axial Seamount Eruption Forecasting Experiment (EFE), a real-time initiative with a reproducible protocol for issuing timestamped, cryptographically hashed volcanic eruption forecasts based on physics-based models and real-time monitoring data, ensuring transparency and rigorous evaluation of forecasting limits.
3. TRIGRSMap: a QGIS plugin for spatio-temporal rainfall-induced landslide susceptibility mapping
Core Problem: The need for enhanced spatial and temporal capabilities in assessing rainfall-induced landslide susceptibility, and the limitations of traditional homogeneous soil parameter assumptions.
Key Innovation: Introduces TRIGRSMap, a novel QGIS plugin that integrates the TRIGRS model with enhanced spatio-temporal capabilities. It demonstrates improved accuracy (40% higher) by using heterogeneous soil parameters and incorporating rainfall variability, providing a more realistic representation of landslide susceptibility zones.
4. Quantifying landslide strain localization phenomena using tensor analysis of multi-temporal lidar data
Core Problem: A fundamental understanding of landslide evolution requires characterizing how deformation localizes within the sliding mass, as traditional analysis often assumes uniform movement.
Key Innovation: Presents a methodology using strain tensor analysis applied to high-resolution displacement fields from multi-temporal UAV-lidar and SfM data to quantify intricate patterns of surface deformation. It computes divergence, gradient, and curl fields to reveal unique kinematic signatures and strain localization behavior for translational and rotational landslides, advancing the mechanistic understanding of slope instability.
5. A regional-scale framework for landslide susceptibility and risk management: application to urban areas with significant cultural heritage in Sicily
Core Problem: The need for a flexible, regional-scale methodology for landslide susceptibility assessment and risk management, particularly for urban areas with cultural heritage, given widespread slope instability.
Key Innovation: Proposes a heuristic methodology for regional-scale landslide susceptibility and risk management. It combines geomorphological, geological, geotechnical, and anthropogenic factors for a susceptibility index, integrates it with an exposure index, and uses a comparison matrix to prioritize interventions and assign risk management categories for urban areas, providing a practical decision-support tool.
6. Evaluation of the regional loess landslides susceptibility based on three-dimensional real scene models
Core Problem: Traditional two-dimensional remote sensing images have shortcomings in accurately and quickly identifying regional loess landslides and evaluating their susceptibility.
Key Innovation: Utilized a three-dimensional real scene model constructed by UAV tilt photogrammetry to accurately and quickly identify loess landslides and perform susceptibility evaluation and zoning using a subjective and objective combination weighting method, achieving high evaluation accuracy (AUC = 0.736) and providing a new approach for disaster prevention.
7. Dynamic response of slopes subject to the coupled action of earthquakes and groundwater
Core Problem: Understanding the complex dynamic response and failure mechanisms of slopes under the combined action of earthquakes and groundwater, which is more intricate than considering them in isolation.
Key Innovation: Investigated the dynamic response of slopes with varying heights and angles to coupled earthquake and groundwater action, revealing non-continuous changes in acceleration amplification factors, the influence of slope angle on AAF distribution, and distinct responses under pulse-like ground motions, contributing to a deeper understanding of slope stability.
8. Analysis and management of geo-hazards through integrated approaches
Core Problem: Landslide studies face challenges due to their multidisciplinary nature, requiring more integrated approaches to advance research and improve the management of affected territories.
Key Innovation: Proposed integrated 'vertical' and 'horizontal' paths for landslide research. The vertical path connects geological structure with landslides across scales and disciplines, while the horizontal path investigates common kinematic characteristics. These approaches enhance landslide hazard assessment and risk management, supporting better zoning, prediction, and mitigation.
9. Enhanced Earthquake Rupture Jumping Distance Across Step‐Overs With Pull‐Apart Basin: Insights From Observation‐Constrained 3‐D Dynamic Rupture Simulations
Core Problem: The effect of pull-apart basins (PABs) in determining earthquake rupture behaviors across step-overs along strike-slip faults remains obscure.
Key Innovation: Showed through 3D dynamic rupture simulations that pull-apart basins greatly enhance rupture jumping capability across step-overs, especially for deep basins, by reducing critical nucleation length and promoting rupture on the second fault, improving understanding of seismic hazards.
10. A divide-and-conquer strategy for fast elastodynamic simulation of earthquakes and aseismic slip on fault networks
Core Problem: Simulating long-term, fully dynamic sequences of earthquakes and aseismic slip (SEAS) on geometrically complex fault networks remains computationally demanding, often necessitating quasi-dynamic approximations.
Key Innovation: An efficient numerical framework for fully elastodynamic SEAS simulations that uses a divide-and-conquer strategy. It treats self-effects and fault-to-fault interactions separately with tailored boundary integral formulations (non-replicating spectral and H-matrices), combined with selective H-matrix compression and physics-informed history truncation, achieving significant speedup and memory reduction.
11. Phys-Diff: A Physics-Inspired Latent Diffusion Model for Tropical Cyclone Forecasting
Core Problem: Deep learning methods for Tropical Cyclone (TC) forecasting often neglect physical relationships between TC attributes, leading to predictions that lack physical consistency.
Key Innovation: Phys-Diff, a physics-inspired latent diffusion model that disentangles latent features into task-specific components (trajectory, pressure, wind speed) and uses cross-task attention to embed physically consistent dependencies, integrating multimodal data for enhanced forecasting performance.
12. Improved MambdaBDA Framework for Robust Building Damage Assessment Across Disaster Domains
Core Problem: Reliable post-disaster building damage assessment (BDA) from satellite imagery is hindered by severe class imbalance, background clutter, and domain shift across different disaster types and geographies.
Key Innovation: Enhancements to the MambaBDA framework, including Focal Loss for class imbalance, Attention Gates for context suppression, and an Alignment Module for spatial warping of pre-event features. These improvements yield consistent performance gains (0.8% to 5% in-domain, up to 27% cross-dataset) and enhance generalization capability for robust building damage assessment across diverse disaster domains.
13. Cross-sphere Coupling and Source Inversion of Ionospheric Disturbances Associated with the 2025 Myanmar Strike-slip Earthquake from BeiDou GEO and Multi-GNSS Observations
Core Problem: Understanding the spatiotemporal evolution and physical mechanisms of pre-seismic and co-seismic ionospheric anomalies associated with earthquakes to enhance monitoring capabilities.
Key Innovation: Utilizes BeiDou GEO and multi-GNSS data to identify a negative Total Electron Content (TEC) anomaly three days pre-earthquake, extract Coherent Ionospheric Disturbance (CID) signals, and proposes a spatial density-weighted method for locating disturbance sources, providing observational constraints on Lithosphere-Atmosphere-Ionosphere Coupling (LAIC) processes.
14. Toward Trustworthy Earthquake Catalogs in the Era of Automated Detection: A Probabilistic Framework for Robust Earthquake Location
Core Problem: Automated earthquake detection methods introduce false detections, mis-associated arrivals, and poorly constrained events, leading to untrustworthy earthquake catalogs and a lack of rigorous uncertainty quantification.
Key Innovation: A fully probabilistic earthquake location framework that jointly infers hypocenters, origin times, noise scales, and contamination levels within a unified Bayesian formulation, using a two-level hierarchical strategy (Student-t scale-mixture and explicit two-component contamination model) and a neural-network travel-time surrogate for scalable, robust, and statistically trustworthy earthquake catalog generation.
15. Insights into evolution of rockfalls on a high-steep slope using UAV photogrammetry and cone complementary-based 3D-DDA
Core Problem: Accurately capturing nonlinear contact interactions in rockfall simulations and constructing precise numerical models of complex slope terrains are challenging for predicting rockfall evolution and impact zones.
Key Innovation: Reformulation of 3D-DDA using cone complementary theory for better nonlinear contact interaction capture, combined with UAV photogrammetry for accurate terrain modeling, enabling enhanced prediction of rockfall trajectories, impact zones, and deposition sites on high-steep slopes.
16. Advancing glacial lake hazard and risk assessment in Bhutan through hydrodynamic flood mapping and exposure analysis
Core Problem: Traditional glacial lake outburst flood (GLOF) hazard and risk assessments in Bhutan have limited consideration of downstream exposure and vulnerability, leading to less robust prioritization of mitigation efforts.
Key Innovation: Models hypothetical GLOF scenarios for numerous glacial lakes, explicitly accounting for downstream impacts using depth-velocity outputs and community vulnerability. This approach identifies previously unrecognized high-hazard lakes and high-risk local government administrative units (LGUs), providing a more robust basis for GLOF preparedness and risk mitigation.
17. Urban flood risk projection under climate change: an approach based on explainable and optimized ensemble learning
Core Problem: Urban floods occur more frequently and respond nonlinearly under climate change, complicating forward-looking risk projection.
Key Innovation: Proposes an explainable and optimized ensemble-learning framework for climate-conditioned urban flood risk projection, integrating 12 risk indicators to improve predictive robustness while retaining interpretable driver attribution.
18. Exploring latent heat flux anomalies for seismic activity detection: insights from RST methodology and case studies
Core Problem: Existing studies on pre-seismic thermal anomalies (LHF) for earthquake prediction are limited by insufficient spatiotemporal resolution and poor data consistency.
Key Innovation: Integrated multi-resolution LHF products with the Robust Satellite Technique (RST) to systematically analyze pre-seismic anomalies for four major earthquakes in China, validating the effectiveness of this combination for detecting potential seismic precursors and providing new observational evidence for remote sensing-based earthquake monitoring.
19. Site specific seismic hazard analysis of Mewa Khola power house site using density approach
Core Problem: Accurate assessment of seismic hazard for critical infrastructure like the Mewa Khola Hydroelectric Project Powerhouse is crucial, but national building codes may underestimate seismic demands for specific locations.
Key Innovation: Conducted a site-specific Probabilistic Seismic Hazard Analysis (PSHA) for the Mewa Khola Powerhouse, incorporating varying earthquake densities based on historical records using the kernel estimation method, revealing that the site's Design Basis Earthquake PGA (0.55 g) exceeds national building code values.
20. Fatal earthquakes in Türkiye and real-time casualty estimates after the 2011 M7.1 Van earthquake
Core Problem: Türkiye faces a high risk of earthquake fatalities, and there is a critical need for accurate and timely real-time casualty estimates to improve disaster response, as current reporting delays are unacceptably long.
Key Innovation: Analyzed historical earthquake fatalities in Türkiye and demonstrated the capability of the QLARM program to provide accurate real-time casualty estimates and rapid alerts (within 30-59 minutes) for major earthquakes, highlighting areas for improvement in data accuracy and reporting speed.
21. Joint use of field investigations and 2D simulations of rockfall trajectories to calibrate surface parameters and generate local propagation maps
Core Problem: Accurately calibrating surface parameters for rockfall trajectory simulations and generating reliable local propagation maps are challenging but essential for effective rockfall management.
Key Innovation: Proposed a two-phase method combining field investigations with 2D simulations to calibrate surface parameters, followed by 3D simulations to generate local rockfall propagation maps. Demonstrated the method's efficiency and practicability for rockfall runout zoning on a real case study.
22. SPT-based liquefaction hazard map using GIS: a case study in Yıldırım district of Bursa, Turkey
Core Problem: Assessing liquefaction potential and producing reliable hazard maps, especially considering groundwater level variability and data uncertainties, is crucial for urban planning and disaster risk reduction.
Key Innovation: Assessed liquefaction potential using SPT data and laboratory tests for a design earthquake scenario, producing GIS-based liquefaction hazard maps under varying groundwater conditions. Highlighted the significant increase in risk with rising groundwater and the importance of considering data density and distribution uncertainties for reliable results.
23. Experimental investigation on a setback distance model for independent foundation overburden sites under normal-fault bedrock dislocation
Core Problem: Ground deformation from strong earthquake fault ruptures causes significant damage to building structures, necessitating a scientific basis to understand failure mechanisms and determine safe setback distances.
Key Innovation: Conducted large-scale physical model tests to simulate bedrock dislocation, exploring the influence of fault dip angle, soil properties, and overburden thickness on earthquake-induced ground rupture and foundation failure. Identified a three-stage progressive failure process and refined the setback distance model, providing a scientific basis for building safety in bedrock dislocation sites.
24. Pervasive distribution of extension-dominated landforms: implications for the kinematics of glacier-free rock avalanches in the Tibetan Plateau highlands
Core Problem: Understanding the emplacement mechanism and dynamics of large Holocene rock avalanches in non-glacial areas, particularly the genetic mechanism of longitudinal ridges, which differ from those in glacial regions.
Key Innovation: Investigated a large Holocene rock avalanche in the Tibetan Plateau, identifying extensive extension-dominated landforms and proposing an extension-dominated propagation process enhanced by self-excited vibration and frozen layer lubrication, explaining the formation of longitudinal ridges.
25. Seismic vulnerability functions accounting for losses on non-structural elements and contents for regional seismic risk assessment
Core Problem: The oversimplification of current seismic vulnerability models, which often neglect the response at the floor level and the distribution of non-structural elements and contents, leading to inaccurate loss estimations for regional seismic risk assessment.
Key Innovation: Development of seismic vulnerability functions for RC infilled frames using a storey-loss-based estimation approach (generalized Storey-Loss Functions) and Modified Cloud Analysis, explicitly accounting for losses on non-structural elements and contents, improving regional seismic risk assessment.
26. Probabilistic and deterministic fault displacement hazard assessment of the Indian region
Core Problem: Conventional deterministic approaches for fault displacement estimation neglect site-specific seismicity, rupture complexity, and structural location, leading to uncertainty in earthquake-resistant design.
Key Innovation: Develops a combined probabilistic and deterministic framework (PFDHA and DFDHA) for fault displacement hazard assessment in the Indian region, incorporating logic trees for epistemic uncertainties, generating hazard maps, and recommending design fault displacement values for seismic zones.
27. Assessing risk of pillar instabilities in abandoned mine workings through probabilistic modelling: A case study in the West Midlands, UK
Core Problem: Assessing the long-term stability of abandoned mine workings and the associated risks of post-mining subsidence is challenging due to complex, time-dependent rock mass behavior, computational demands of existing models, and uncertainties in rock properties and mine geometries.
Key Innovation: Develops simplified yet practical analytical models integrated with probabilistic methods (Monte Carlo simulations) for rapid stability assessments of abandoned mine workings, demonstrating how inadequate pillar design and moisture-sensitive roof strata contribute to pillar foundation failure and surface subsidence, providing guidance for risk mitigation.
28. Evolution characteristics of fissure structure and permeability of undisturbed highly expansive soils in cold regions under the full range of seasonal variation
Core Problem: Understanding how seasonal climatic variations (wetting-drying-freeze-thaw cycles) degrade the microscopic fissure structure and macroscopic hydraulic properties of undisturbed highly expansive soils, which increases the risk of shallow landslides.
Key Innovation: Conducted WDFT cycle tests, digital photography, SEM, and X-ray CT to show that seasonal variation promotes fissure propagation, accelerating water infiltration. The study found that a stable fissure network forms after five WDFT cycles and that interconnected fissure quantity and volume are primary factors for soil permeability.
29. Refined constitutive modelling of widely graded granular soils incorporating fractional particle breakage
Core Problem: Accurate characterization of gradation evolution and its influence on the critical state line (CSL) is essential but challenging for modeling the mechanical behavior of widely graded granular soils.
Key Innovation: Developed a refined constitutive model for widely graded granular soils by formulating a fractional breakage evolution model, establishing a gradation-dependent CSL based on the breakage-packing concept, and proposing a state-dependent dilatancy equation using fractional-order plasticity theory, demonstrating robustness in predicting stress-strain relationships and gradation evolution.
30. A depth-averaged method modeling three-dimensional submarine landslides caused by earthquakes
Core Problem: Conventional methods for modeling earthquake-triggered submarine landslides simplify seismic loading and neglect strain-softening behavior of marine sediments, limiting their ability to reproduce the entire landslide process.
Key Innovation: Development of a new depth-averaged method that incorporates 3D seismic accelerations and curvature effects, capturing shear band propagation from initiation to post-failure runout, offering a computationally efficient tool for assessing large-scale submarine slides.
31. Bayesian optimized random forest models for predicting permanent displacement considering fault types and pulse-like ground motion effects
Core Problem: Many existing predictive displacement models for earthquake-induced landslides do not explicitly account for fault type, ground motion orientation, or the probability of pulse-like ground motion (PLGM), leading to incomplete assessments of permanent displacement.
Key Innovation: Develops new predictive displacement models using a Bayesian-optimized random forest regressor that explicitly incorporates pulse probability and slope-fault orientation, enabling direction-dependent displacement estimates within a unified framework for regional seismic landslide hazard categorization.
32. The origin and characteristics of seismic swarms in the Abu Dabbab district, Egypt's Eastern Desert, deduced from the analysis of seismograms
Core Problem: Understanding the mechanisms and origins of recurring earthquake swarms in the Abu Dabbab district is crucial for accurate seismic hazard assessments.
Key Innovation: Analysis of 4058 earthquakes identified distinct types, including long-period (LP) and very long-period (VLP) signals, suggesting elastic ground deformation driven by pressure within a magma column. The study highlights the combined influence of regional rift-related extension, dike intrusions, and crustal heterogeneity, improving seismic hazard assessment accuracy.
33. Cyclic liquefaction resistance of saturated clayey sands under asymmetrical cyclic loading conditions
Core Problem: There is a lack of data on the impact of clay inclusion on the liquefaction susceptibility of sloped grounds under asymmetrical seismic loading, which can lead to severe consequences.
Key Innovation: A systematic experimental study evaluated the liquefaction susceptibility of clayey Toyoura sands, identifying various failure modes and demonstrating that clay inclusion significantly reduces initial static shear stress correction factors (Kα) and threshold α (αth), which must be considered in practical applications. Unified linear correlations between CRR and initial state parameter (ψ) were also established.
34. Temperature dependence of broadband seismometer sensitivity
Core Problem: The temperature dependence of broadband seismometer sensitivity is a limiting factor for calibration uncertainty and observation accuracy in global geophysics networks, and accurate measurement systems for this coefficient have not been established.
Key Innovation: Development of a system using a triaxial vibration exciter combined with a thermostatic chamber to accurately measure the temperature coefficient of seismometer sensitivity, finding a coefficient of (0.11±0.01) %/°C and stable relative frequency response below 1 Hz, which is useful for evaluating measurement accuracy.
35. Open-Vocabulary vs Supervised Learning Methods for Post-Disaster Visual Scene Understanding
Core Problem: Automated interpretation of aerial imagery for large-scale post-disaster damage assessment is challenging due to clutter, visual variability, strong cross-event domain shift, and the reliance of supervised approaches on costly, task-specific annotations with limited coverage.
Key Innovation: A comparative evaluation of open-vocabulary and supervised learning models for post-disaster scene understanding (semantic segmentation and object detection), demonstrating that supervised training remains the most reliable approach when labels are fixed and available, particularly for small objects and fine boundary delineation.
36. Cryo-Bench: Benchmarking Foundation Models for Cryosphere Applications
Core Problem: There is a lack of suitable evaluation datasets for benchmarking Geo-Foundation Models (GFMs) in Cryosphere applications, limiting understanding of their performance and optimal usage strategies for monitoring critical components like glaciers and glacial lakes.
Key Innovation: Introduces Cryo-Bench, a comprehensive benchmark compiled to evaluate GFM performance across key Cryospheric components (debris-covered glaciers, glacial lakes, sea ice, and calving fronts). It evaluates 14 GFMs and provides recommendations for optimal usage strategies, such as encoder fine-tuning with hyperparameter optimization.
37. Probabilistic Retrofitting of Learned Simulators
Core Problem: Dominant approaches for modeling Partial Differential Equations (PDEs) rely on deterministic predictions, failing to capture the inherent chaos and uncertainty in many physical systems, and training probabilistic models from scratch is computationally expensive.
Key Innovation: Proposes a training-efficient, architecture-agnostic strategy to transform pre-trained deterministic PDE models into probabilistic ones via retrofitting with the Continuous Ranked Probability Score (CRPS). This method significantly reduces rollout CRPS and improves variance-normalised RMSE, demonstrating that probabilistic PDE modeling can be unlocked from existing deterministic backbones with modest additional training cost.
38. CASCADE: Cross-scale Advective Super-resolution with Climate Assimilation and Downscaling Evolution
Core Problem: Super-resolution of geophysical fields, especially for extreme events, faces unique challenges as fine-scale structures must respect physical dynamics, conserve mass/energy, and evolve coherently in time, which traditional image enhancement methods struggle to achieve.
Key Innovation: CASCADE, a framework that reframes spatiotemporal super-resolution as an explicit transport process across scales, using learned flow-conditioned velocity fields and an assimilation-style innovation step, producing temporally coherent, physically consistent, and mass-conserving reconstructions for severe convective storms.
39. Learning with the Nash-Sutcliffe loss
Core Problem: The Nash-Sutcliffe efficiency (NSE), while widely used for evaluating time series forecasts in fields like hydrology, lacks a decision-theoretic foundation for its use as a loss function for model estimation, leading to implicit assumptions about data generation.
Key Innovation: Proves that the Nash-Sutcliffe loss ($L_{\text{NS}} = 1 - \text{NSE}$) is strictly consistent for an elicitable and identifiable multi-dimensional functional, establishing a decision-theoretic foundation for NSE-based model estimation and forecast evaluation, and introduces Nash-Sutcliffe linear regression for multi-dimensional time series forecasting.
40. OSDM-MReg: Multimodal Image Registration based One Step Diffusion Model
Core Problem: Existing multimodal remote sensing image registration methods struggle to extract modality-invariant features and bridge large nonlinear radiometric differences between images (e.g., SAR and optical), hindering effective data fusion.
Key Innovation: Proposes OSDM-MReg, a novel framework that uses a one-step unaligned target-guided conditional diffusion model (UTGOS-CDM) for rapid image-to-image translation into a unified representation domain, combined with a multimodal multiscale registration network (MM-Reg) to achieve superior registration accuracy for multimodal remote sensing images.
41. RED-DiffEq: Regularization by denoising diffusion models for solving inverse PDE problems with application to full waveform inversion
Core Problem: Partial differential equation (PDE)-governed inverse problems, such as full waveform inversion in geophysics, are challenging due to nonlinearity, ill-posedness, and sensitivity to noise, hindering accurate reconstruction of subsurface models.
Key Innovation: RED-DiffEq, a new computational framework, integrates physics-driven inversion and data-driven learning by leveraging pretrained diffusion models as a regularization mechanism for PDE-governed inverse problems. Applied to full waveform inversion, it shows enhanced accuracy, robustness, and generalization ability for reconstructing high-resolution subsurface velocity models.
42. CASCADE-3D: A GUI-Driven Framework for Automated 3D Building Model Reconstruction
Core Problem: The need for rapid and accurate generation of geospatial data and 3D building features to support multipurpose land management services, particularly for automated reconstruction of detailed 3D building models.
Key Innovation: Presenting CASCADE-3D, a GUI-driven framework that integrates advanced deep-learning (SAM, DGCNN) for automated 3D building model reconstruction (LOD1 and LOD2), including building outline detection, point cloud classification, and precise roof structure extraction, achieving high accuracy (RMSE 0.36 m) and exporting in CityJSON format.
43. Three-Branch Multiscale Abundance Feature Fusion Network for Hyperspectral Image Change Detection
Core Problem: Hyperspectral image change detection (HCD) is hindered by mixed pixels leading to coarse results, ineffective fusion of pixel-level and subpixel-level features, and unsatisfactory performance for minority classes due to class imbalance.
Key Innovation: Proposing a three-branch multiscale abundance feature fusion network (TMAFN) that introduces superpixel-constrained abundance maps, develops a spatial–channel–spatial module for fine-grained feature extraction, and designs a multihead cross attention mechanism with multiscale residual fusion to enhance detection of subtle and imbalanced change categories.
44. Boosting MonoDepth With Foundation Models and Edge Laplace Cross-Entropy in Remote Sensing
Core Problem: Monocular depth estimation from single remote sensing images is challenging due to blurriness and overlap at object edges, and identifying the most suitable architecture for practical applications remains difficult.
Key Innovation: Harnessing vision foundation models for rich feature extraction and integrating an edge Laplacian cross-entropy loss to mitigate blurriness and overlap at object edges by transforming unimodal depth modeling into a multimodal approach, achieving superior depth estimation results on remote sensing datasets.
45. Multimodel Comparative Assessment and Optimal Geodetector-Driven Analysis of Aeolian Desertification in the Horqin Sandy Land
Core Problem: Precise monitoring of aeolian desertification (AD) dynamics and understanding its natural-anthropogenic driving mechanisms are crucial for maintaining ecological stability in critical regions like the Horqin Sandy Land.
Key Innovation: Innovatively integrated multisource remote sensing parameters via Google Earth Engine to compare three feature space models and five machine learning algorithms for high-precision AD monitoring, identifying Random Forest as optimal. It also applied Optimal Parameter Geographic Detector (OPGD) to quantify natural-anthropogenic driving mechanisms, revealing soil texture as a strong single driver and significant interactive effects.
46. Global mapping of lake-terminating glaciers
Core Problem: A consistent global inventory of lake-terminating glaciers has been lacking, despite their importance in accelerating melt, velocity, and contributing to cryospheric hazards.
Key Innovation: Presents a global inventory of lake-terminating glaciers, differentiating between three categories based on glacier-lake contact. This dataset, integrated into RGI 7.1, identifies 1.4% of global glaciers as lake-terminating, accounting for 11.4% of global glacier area, providing a crucial resource for cryospheric hazard assessment.
47. Conditional diffusion models for downscaling and bias correction of Earth system model precipitation
Core Problem: Earth System Models (ESMs) struggle to accurately simulate high-resolution precipitation, particularly for extreme events like heavy rainfall and flooding, due to issues with resolving small-scale dynamics and inherent biases, while existing statistical and deep learning methods have limitations in improving spatial structure or stability.
Key Innovation: A novel machine learning framework using conditional diffusion models for simultaneous bias correction and downscaling of ESM precipitation, which maps observational and ESM data to a shared unbiased embedding space, ensuring statistical fidelity, preserving spatial patterns, and outperforms existing methods, especially for extreme events.
48. Evolution of watershed resilience research in China: a bibliometric analysis
Core Problem: Climate change and intensified human activities have exacerbated watershed risks in China, and existing reviews lack systematic integration of resilience enhancement research.
Key Innovation: A bibliometric analysis of watershed resilience research in China to systematically integrate existing knowledge and identify research trends and gaps.
49. Research on flood prediction based on hybrid AI model VMD_BiLSTM: a case study in BeiJing flood, China
Core Problem: Traditional machine learning models struggle to capture the nonlinear dynamics of rapid water level fluctuations caused by intense rainfall, leading to reduced predictive performance for flood prevention.
Key Innovation: Proposed a hybrid VMD-BiLSTM model that decomposes complex time series into stable patterns, significantly improving flood water level prediction accuracy (R2 of 0.940, RMSE of 0.008 m) compared to traditional models, offering practical value for disaster prevention.
50. Deep learning based flood segmentation: evaluating EfficientNet and ResNet for UNet, SegNet and DeepLabV3+
Core Problem: Conventional flood mapping methods suffer from high computational demands, coarse spatial resolution, and limited real-time adaptability, hindering effective disaster response.
Key Innovation: Systematically compared deep learning segmentation architectures (UNet, SegNet, DeepLabV3+) with EfficientNet/ResNet backbones for flood delineation, finding UNet with EfficientNet backbone to be the most efficient and accurate (accuracy 0.9705), offering a scalable solution for near real-time disaster management.
51. Spatial multi-criteria evaluation approach for flood hazard and vulnerability assessment of Jammu City Region, India
Core Problem: The Jammu City Region faces significant flood threats from the Tawi River due to altered river course, encroachments, and rapid urban expansion, necessitating a comprehensive assessment of flood hazard and vulnerability.
Key Innovation: Investigated flood hazard and vulnerability using an integrated spatial multi-criteria evaluation approach with the analytic hierarchy process, identifying nearly one-third of the region as flood-susceptible and specific high-vulnerability zones, and recommending a river-centric development approach for mitigation.
52. Toward an accurate assessment of flood resilience enhancement: comprehensive comparisons using a novel performance-based metric integrating resilience-enhancing effects
Core Problem: Existing performance-based flood resilience metrics neglect the enhanced ability of the physical environment to manage flooding after green infrastructure (GI) implementation, hindering accurate assessment of resilience-enhancing effects.
Key Innovation: Proposed a novel performance-based metric that accounts for the resilience-enhancing effects of Green Infrastructure (GI) and conducted comparative analyses, demonstrating that GI enhances system performance and reduces impact duration, especially in severely inundated areas, providing a more accurate assessment of flood resilience.
53. Land uplift induced by groundwater rise in thick deep soft clay strata and its impact on metro tunnels: a case study from Tianjin in China
Core Problem: The impact of groundwater recovery-induced ground rebound on underground infrastructure, such as metro tunnels, is not well understood.
Key Innovation: Investigated spatiotemporal groundwater rise and ground deformation using SBAS-InSAR and hydrogeological observations, revealing a strong correlation and potential impacts on metro tunnels (uplift, leakage, differential deformation). Provided insights for managing groundwater-induced ground rebound and associated disasters.
54. Fire following earthquake: a comprehensive review of ignition, spread, and risk assessment models
Core Problem: Fire following earthquake (FFE) poses significant risks, and a comprehensive understanding of its initiation mechanisms, spread dynamics, and effective risk assessment techniques is needed to develop mitigation strategies.
Key Innovation: Provided a comprehensive review of FFE research, synthesizing studies on initiation, spread, and risk assessment models. Identified knowledge gaps and outlined future research directions to contribute to the development of effective FFE mitigation strategies.
55. Experimental study of the contact area effect on normal deformability of rock joints
Core Problem: The normal deformability of rock joints is crucial for the mechanical behavior of jointed rock masses (e.g., rock slopes, underground structures), but the impact of contact area on normal stiffness remains a debated topic.
Key Innovation: Conducted extensive laboratory experiments on natural rock joints with varying roughness and materials to investigate the impact of contact area on normal stiffness. Found that contact area, controlled by normal stress, roughness, and material type, significantly influences joint closure, normal stiffness, and stress transmission, providing a nonlinear correlation for contact ratio.
56. Performance of Tunnel Flexible Protection Mesh Under Rockburst Impact Loading: A Numerical Simulation Study
Core Problem: Protecting deep tunnels from rockbursts is challenging, requiring effective flexible protection systems that can absorb significant impact energy.
Key Innovation: Proposed a reliable numerical modeling method for flexible protection mesh under rockburst impact. Revealed dynamic response characteristics and optimized design factors (bolt spacing, plate size, bolt pattern) for enhanced protection, showing impact energy is mainly absorbed by plastic deformation of the mesh.
57. Deep Cement Mixing in Geotechnical Engineering: Applications, Developments, and Emerging Trends
Core Problem: The need for a comprehensive synthesis of the latest developments, applications, and future trends in Deep Cement Mixing (DCM) as a ground improvement technique for soft soils.
Key Innovation: A comprehensive review synthesizing the latest developments in DCM, covering design principles, construction methodologies, engineering applications, quality control, performance monitoring, and recent progress in experimental and numerical modeling, providing insights for optimization.
58. Combined effects of soil internal and external forces on aggregate fragments detachment and displacement during splash erosion
Core Problem: It is unclear how the combined effects of soil internal forces (SIFs) and external raindrop impact forces influence the detachment and displacement characteristics of aggregate fragments during splash erosion, which is crucial for effective soil erosion prevention.
Key Innovation: This study conducted simulated rainfall experiments using electrolyte solutions to quantitatively regulate SIFs and varying rainfall heights for raindrop impact, demonstrating that an electrolyte concentration of 10−2 mol L−1 represents a critical point affecting SIFs and splash erosion, and that both internal and external forces control aggregate fragment detachment and displacement, providing new insights into splash erosion mechanisms.
59. Landscape-scale vegetation cover shapes microbial community assembly and functional potential via Erosion-driven nutrient patterns in Alpine Rivers
Core Problem: The impact of soil erosion, strongly influenced by vegetation cover on mountain slopes, on microbial communities (planktonic and benthic) and their functional potential in alpine river ecosystems, particularly on the Qinghai–Tibetan Plateau, remains largely unexplored.
Key Innovation: This study analyzed planktonic and benthic microbiomes under different normalized difference vegetation index (NDVI) conditions, revealing that lower NDVI (indicating higher erosion) correlates with elevated nutrient concentrations in riparian sediments and higher benthic microbial diversity, and that NDVI regulates carbon/nitrogen cycling through nutrient-driven bacterial processes and functional gene coupling, elucidating structural and functional dynamics of eroded alpine river microbial communities.
60. Mechanical responses during loess tunnel construction resulting from groundwater level changes due to injection well leakage
Core Problem: Groundwater perturbations, specifically from injection well leakage, disrupt seepage-stress equilibrium during loess tunnel excavation, leading to instability and increased failure risk.
Key Innovation: Developed a fluid-solid coupling model and conducted numerical simulations and model tests to quantify stress-displacement responses. Findings include inclined groundwater tables, heterogeneous radial stress redistribution, stress asymmetry in lining, and lateral shift of surface settlement, providing insights into tunnel construction under varying groundwater conditions.
61. A computational framework for evaluating three-dimensional segmental tunnel–ground interaction under long-term joint leakage
Core Problem: The long-term impacts of joint leakage in segmental tunnels on the ground-tunnel system, influenced by gasket ageing, bi-directional hydraulic conditions, and coupled with joint deformations, are insufficiently explored.
Key Innovation: Proposed a 3D FEM-based computational framework incorporating theoretical models for hydro-mechanical coupled behaviors at gasketed joints. The framework demonstrated that localized leakage significantly increases bending moments, reduces axial forces, and causes notable transverse settlement and widespread groundwater drawdown, with potential for reversed structural responses under extreme conditions.
62. The impact of a mega-flood event on the water quality of the southern Murray-Darling Basin, Australia
Core Problem: Uncertainty regarding the impacts of increasingly frequent and severe extreme floods, driven by climate change, on water quality in river systems like the Murray-Darling Basin.
Key Innovation: Investigated the dynamics of TN, TP, and DOC during six major flow events, including a 2022–2023 'mega-flood', using statistical and hysteresis analysis. Found the mega-flood accounted for over 30% of total flow and significant nutrient yield, with distinct counter-clockwise hysteresis indicating delayed release from floodplains, demonstrating that mega-floods lead to proportionally larger nutrient loads and prolonged water-quality degradation.
63. Probabilistic analysis of slopes designed by the partial material factor approach considering Gaussian copula-based cross-correlated random fields
Core Problem: Conventional slope design methods often neglect soil spatial variability and non-linear cross-correlations between soil properties, leading to inconsistent reliability levels and potentially unconservative designs.
Key Innovation: Combines copula theory with the Random Finite Element Method to construct cross-correlated random fields, capturing non-linear dependencies between soil parameters for probabilistic slope design, demonstrating that neglecting non-linear dependence can lead to significantly underestimated failure probabilities.
64. A centrifuge model test on the influence of pile-soil structure on soft soil seismic response based on Hilbert Huang Transform
Core Problem: Pile-soil interaction in soft soil sites during seismic loading is often simplified or overlooked in seismic design, leading to vulnerability of infrastructure to strong earthquakes and associated secondary hazards.
Key Innovation: A dynamic centrifuge test and numerical simulation revealed that pile foundations significantly alter local seismic responses in soft soils, leading to distinct energy accumulation and release patterns, enhanced low-frequency amplification, and a fundamental difference in dynamic response mechanisms between pile-influenced and free-field zones.
65. An example of regional seismicity recovery on June 22, 2020, using waveform cross-correlation
Core Problem: The International Monitoring System (IMS) for seismic events may have detection thresholds that are not low enough to consistently detect all significant events, potentially missing weapon-sized explosions or other seismic occurrences, especially in continental areas.
Key Innovation: Demonstrates the effectiveness of waveform cross-correlation (WCC) techniques to significantly lower the detection threshold for seismic events, enabling the recovery and high-quality reporting of events that were previously undetected or unreported by routine processing.
66. MixerCSeg: An Efficient Mixer Architecture for Crack Segmentation via Decoupled Mamba Attention
Core Problem: Existing CNN-, Transformer-, and Mamba-based models for pixel-level crack segmentation fail to fully capture the required spatial or structural information for modeling complex crack patterns efficiently.
Key Innovation: MixerCSeg, an efficient mixer architecture for crack segmentation that integrates CNN-like, Transformer-style, and Mamba-inspired pathways within a single encoder (TransMixer) to capture local textures, global dependencies, and sequential context. It also introduces a Direction-guided Edge Gated Convolution (DEGConv) and a Spatial Refinement Multi-Level Fusion (SRF) module for enhanced structural fidelity and multi-scale detail refinement, achieving state-of-the-art performance with high efficiency.
67. Tri-path DINO: Feature Complementary Learning for Remote Sensing Multi-Class Change Detection
Core Problem: Multi-class change detection (MCD) in remote sensing imagery is constrained by complex scene variations and the scarcity of detailed annotations, hindering fine-grained monitoring.
Key Innovation: Proposes Tri-path DINO, an architecture that uses a three-path complementary feature learning strategy (DINOv3 for coarse features, auxiliary siamese path for fine-grained features, multi-scale attention for contextual information) to rapidly adapt pre-trained foundation models for MCD, achieving optimal performance on damage assessment datasets.
68. DAWN-FM: Data-Aware and Noise-Informed Flow Matching for Solving Inverse Problems
Core Problem: Inverse problems, common in fields like geophysics, are often ill-posed and require robust regularization techniques to estimate parameters from incomplete or noisy observations, especially when data is noisy or incomplete.
Key Innovation: DAWN-FM (Data-AWare and Noise-Informed Flow Matching), a generative framework that integrates a deterministic process to map a reference distribution to the target, incorporating data and noise embedding to explicitly access measured data representations and account for noise, making it robust for inverse problems in fields like geophysics.
69. LD-EnSF: Synergizing Latent Dynamics with Ensemble Score Filters for Fast Data Assimilation with Sparse Observations
Core Problem: Existing score-based data assimilation methods for tracking complex dynamical systems are computationally expensive, especially when dealing with high-dimensional, nonlinear systems and sparse observations, due to the need for full-space simulations.
Key Innovation: Proposes LD-EnSF, a novel score-based data assimilation method that eliminates full-space simulations by evolving dynamics in a compact latent space using improved Latent Dynamics Networks (LDNets) and a history-aware LSTM encoder for sparse observations, achieving significant speedups and maintaining high accuracy.
70. State parameter from CPT under high overburden stress: a nonlinear critical state approach
Core Problem: Traditional methods for estimating the state parameter (ψ) from CPT data assume a linear critical state line, which becomes inaccurate under high overburden stress due to significant particle crushing.
Key Innovation: A refined framework based on Bolton’s relative dilatancy index (IR) that inherently accounts for a nonlinear critical state line and stress-dependent soil behavior, providing an improved method for estimating ψ, particularly under high-stress conditions.
71. Adaptive Superpixel Segmentation-Based Coastline Extraction Method for PolSAR Images
Core Problem: Improving the accuracy of coastline extraction from polarimetric synthetic aperture radar (PolSAR) images, which is challenging due to image complexity.
Key Innovation: Develops an adaptive superpixel segmentation-based method that extracts multiple polarimetric and texture features, determines the optimal number of superpixels, segments them using SLIC, and merges them with a fractal network evolution algorithm (FNEA) to achieve high precision in coastline extraction.
72. Assessing the intensification and impact of a historical storm in a warmer climate
Core Problem: Uncertainty regarding the implications of future climate warming on wind-related impacts from extratropical windstorms in Northern Europe.
Key Innovation: Investigates the response of thermodynamic warming to a historical storm using a convection-permitting numerical weather prediction model within a pseudo-global warming framework, demonstrating systematic intensification of near-surface wind and gust speeds, expanded spatial footprint, and increased cumulative wind exposure, implying enhanced potential for wind-related impacts.
73. Atmospheric and cryospheric observations at the high-altitude Zarafshon River Basin and the Hydrographic Party Glacier (GGP), Tajikistan, 2018–2025
Core Problem: Scarcity of cryospheric and atmospheric observations in Central Asia, despite the integral role of glaciers and snow cover in the regional hydrological cycle and the need to assess climate change impacts.
Key Innovation: Presents a diverse dataset of cryospheric and atmospheric variables (glacier terminus position, snow conditions, ablation, aerial photography, meteorological variables, surface reflectance, snow chemistry, atmospheric aerosols) from the Zarafshon River Basin and GGP, Tajikistan, providing a valuable basis for research on glacier dynamics, snow processes, and atmosphere-cryosphere interactions.
74. Dataset of a 4 km combined seismic and electric streamer survey along the embankment of the Po river in Crescentino
Core Problem: Characterizing river embankments using non-invasive geophysical investigations, particularly demonstrating the applicability of combined seismic and electric streamer systems for long and extensive surveys, and the need for tailored processing methods for large datasets.
Key Innovation: Presents a dataset from a 4 km combined seismic and electric streamer survey along the Po River embankment, demonstrating the efficiency and potential of these systems for investigating river embankments and providing a structured dataset for testing and benchmarking processing and interpretation approaches.
75. Climate impacts from North American boreal forest fires
Core Problem: Quantifying the net climate impacts (warming vs. cooling) of boreal forest fires in Alaska and western Canada, considering multiple interacting factors like greenhouse gas emissions, vegetation recovery, permafrost thaw, and surface albedo changes over a 70-year period.
Key Innovation: Estimation of integrated net radiative forcing metrics for boreal fires (2001-2019), revealing that Alaskan fires generally contribute to net warming (0.35 ± 4.66 W m−2) due to high fuel consumption and permafrost thaw, while Canadian fires contribute to net cooling (−2.88 ± 4.17 W m−2) by increasing snow albedo, and identifying specific landscape characteristics associated with warming vs. cooling impacts.
76. Identifying information voids during weather-related diasters: case studies from the 2024 Europe floods and Florida’s hurricane helene
Core Problem: Information voids are anecdotally linked to misinformation during crises but lack empirical testing, hindering effective disaster risk communication and potentially impacting lives and livelihoods.
Key Innovation: Conceptualized and tested a novel survey tool to empirically detect information voids across four dimensions (quantity, quality, source, channel) during climate emergencies (floods, hurricanes), demonstrating its reliability and sensitivity for generating actionable insights to improve disaster communication.
77. Dynamic Response and Safety Control of Stratified Rock in Tunnels Under Dynamic and Static Loads
Core Problem: Ensuring safe construction of tunnels by drill and blast excavation, which requires clarifying stress and deformation characteristics of stratified rock under dynamic (blasting) and static (in-situ) loads.
Key Innovation: Used similar proportion tests, static deformation model tests, and LS-DYNA numerical simulations to analyze deformation and stress distribution in stratified rock in a deep-buried diversion tunnel. Established a predictive model for stratified rock (peak and secondary equilibrium stress) and determined limiting dynamic and static loads for safety control based on rock inclination.
78. A 2D Fully Coupled, Thermo-hydro-mechanical Modeling for Fault Activation Using the Combined Finite–Discrete Element Method (FDEM)
Core Problem: Investigating the underlying mechanism of fault reactivation induced by cold-fluid injection during geothermal heat extraction.
Key Innovation: Proposed a 2D fully coupled thermo-hydro-mechanical (THM) model within the FDEM framework. Verified its accuracy and robustness, and applied it to simulate cold-fluid injection, demonstrating how initial rock temperature, injected pressure, and thermal expansion coefficient affect fault slip, aperture, and acoustic emission events.
79. Quantitative Analysis of Frequency–Energy Characteristics of Blast-Induced Vibrations in Deep Parallel Tunnels
Core Problem: Quantifying and understanding the frequency-energy characteristics of blast-induced vibrations in deep parallel tunnels is essential for managing their impact and ensuring structural stability.
Key Innovation: Presents a quantitative frequency-energy characterization of blast-induced vibrations in deep parallel tunnels, providing analyzable metrics that support vibration control and construction safety management.
80. Simple Shear Test of Prismatic Rock Specimens and Influence of Specimen Shape on Mechanical Properties of Anisotropic Tuff
Core Problem: There is no standardized method for simple-shear testing of rocks in Japan, and precise identification of mechanical properties for evaluating seismic stability of rock masses and stiffness of basement rocks for seismic response analysis remains elusive.
Key Innovation: Introduced a novel simple-shear test method for prismatic rock specimens using a torsional-shear apparatus. Demonstrated that shorter specimens yield more uniform stress–strain curves and Mohr’s stress circles consistent with failure criteria, confirming the method's effectiveness and suggesting optimal specimen dimensions for evaluating seismic stability.
81. A Simplified Decoupled Computational Framework for Effective Stress Analysis in Geotechnical Engineering
Core Problem: The complexity and computational cost of effective stress analysis in geotechnical engineering, especially for both saturated and unsaturated soils, hindering efficient stability and deformation analysis of critical structures.
Key Innovation: A simplified decoupled computational framework for effective stress analysis, based on seepage lines, capable of directly computing effective stress fields for geotechnical structures (like dams and tailings ponds) with high efficiency, particularly for gentle phreatic lines.
82. Multistage Shear Behavior of Gravel–Sand–Fines Mixtures Based on the RSCS Framework
Core Problem: The variability and method-dependency of shear strength parameters obtained from multistage direct shear tests on gravel-sand-fines mixtures, particularly concerning the influence of displacement-reset procedures.
Key Innovation: Evaluation of three multistage direct shear methods, demonstrating that the degree of peak-strength mobilization is strongly influenced by the displacement-reset procedure and soil composition, emphasizing the need to account for fabric renewal for reliable shear strength parameter application.
83. Basal Stability of Braced Circular Excavations in Anisotropic and Non-homogeneous Clays
Core Problem: The need for a comprehensive understanding and predictive tools for basal heave stability in braced circular excavations, especially considering the complex effects of anisotropic and non-homogeneous strength conditions in soft clays.
Key Innovation: A numerical study using the NGI-ADP anisotropic strength model to investigate basal heave stability in circular excavations, identifying key influencing factors and deriving regression-based equations for quick and reliable estimation of basal stability under complex soil conditions.
84. Climate-induced soil hydro-thermal response on sunny and shady slopes incorporating crack effects
Core Problem: The long-term impact of sunny-shady slope effects on cracked soil slopes under atmospheric conditions, and their influence on hydro-thermal behavior and geotechnical infrastructure durability, is not fully understood.
Key Innovation: Presents a novel numerical approach (extension of Ev-CRACK model) integrating solar radiation distribution and exponential attenuation along crack surfaces to comprehensively analyze the hydro-thermal response of cracked soil slopes, demonstrating how cracks and slope orientation influence temperature fluctuations and moisture loss.
85. Role of hydrate cementation habits on the geomechanical behavior of hydrate bearing sediments
Core Problem: The influence of hydrate pore morphology on the large-strain response and cyclic deformation of hydrate-bearing sediments, crucial for reliable geomechanical assessment during gas production, is poorly constrained.
Key Innovation: Provides the first systematic experimental comparison of pore-filling and cemented hydrate morphologies under monotonic and slow cyclic loading, demonstrating that hydrate morphology controls peak strength, stiffness degradation, dilation, and long-term mechanical stability, and enhances a hypoplastic constitutive model to improve prediction accuracy.
86. CNN‐Based Retrieval of 3D Cloud Structures Solely From Geostationary Satellite Imagery
Core Problem: Operational cloud vertical structure (CVS) products depend on active observation or reanalysis fields, limiting high-frequency monitoring, and there is a need for satellite-only methods.
Key Innovation: Proposed a lightweight, satellite-only CNN model that reconstructs volumetric cloud masks from geostationary multispectral imagery, achieving strong performance and enabling near-real-time monitoring of rapidly evolving events like typhoon cloud structures.
87. Towards Data-driven Nitrogen Estimation in Wheat Fields using Multispectral Images
Core Problem: Accurate Targeted Spraying and Fertilization (TSF) in agriculture is challenging due to spatio-temporal variability and external factors, requiring data-driven solutions for nitrogen estimation.
Key Innovation: TerrAI, a Neural Network-based solution for TSF that considers spatio-temporal variability across parcels, validated using a real-world remote sensing dataset for nitrogen estimation in wheat fields using multispectral images.
88. Scalable Gaussian process modeling of parametrized spatio-temporal fields
Core Problem: Efficiently and accurately learning continuous representations of parametrized spatio-temporal fields, especially when requiring scalable uncertainty quantification, which is often computationally expensive for traditional Gaussian process methods.
Key Innovation: Introduces a scalable Gaussian process framework with deep product kernels that learns a continuous representation of spatio-temporal fields. It leverages Kronecker algebra for efficient training and posterior variance computation, enabling scalable uncertainty quantification and achieving competitive accuracy with operator learning methods.
89. Improving Full Waveform Inversion in Large Model Era
Core Problem: Existing data-driven Full Waveform Inversion (FWI) methods use small models, leading to overfitting and poor generalization to complex, realistic geological structures despite performing well on synthetic datasets.
Key Innovation: A working recipe that scales a billion-parameter FWI model through coordinated scaling across model capacity, data diversity, and training strategy. This enables a model trained on simple simulated data to generalize remarkably well to challenging and unseen geological benchmarks, significantly narrowing the generalization gap.
90. Dual-space posterior sampling for Bayesian inference in constrained inverse problems
Core Problem: For inverse problems constrained by partial differential equations (e.g., full waveform inversion for Earth's subsurface properties), there is no clear procedure to translate hard physical constraints into prior distributions amenable to existing Bayesian sampling techniques, hindering uncertainty quantification.
Key Innovation: A dual-space posterior sampling method that integrates the alternating direction method of multipliers (ADMM) with Stein variational gradient descent (SVGD) within an augmented Lagrangian formulation. This approach translates hard constraints into penalties, ensuring their exact satisfaction while enabling constrained posterior sampling and providing well-calibrated uncertainty estimates.
91. Station2Radar: query conditioned gaussian splatting for precipitation field
Core Problem: Precipitation forecasting is challenged by heterogeneous data sources (radar, stations, satellites) with limitations such as limited coverage, sparsity, or lack of direct rainfall retrieval, making accurate and dense precipitation field generation difficult.
Key Innovation: Proposes Query-Conditioned Gaussian Splatting (QCGS), a framework that fuses automatic weather station observations with satellite imagery to generate efficient, resolution-flexible precipitation fields in real time, demonstrating significant RMSE improvement over conventional products.
92. High Dynamic Range Imaging Based on an Asymmetric Event-SVE Camera System
Core Problem: Conventional cameras struggle with high dynamic range (HDR) imaging under extreme illumination due to overexposure, limiting reliable perception in challenging environments.
Key Innovation: A hardware-algorithm co-designed HDR imaging system integrating an SVE micro-attenuation camera with an event sensor, featuring a two-stage cross-modal alignment framework and a cross-modal HDR reconstruction network with convolutional fusion and mutual-information regularization, improving highlight recovery, edge fidelity, and robustness.
93. Cross-Scale Pansharpening via ScaleFormer and the PanScale Benchmark
Core Problem: Existing pansharpening methods struggle with generalization to real-world, high-resolution, cross-scale scenarios due to limited evaluation settings.
Key Innovation: Introduction of PanScale, a large-scale cross-scale pansharpening dataset and benchmark, and ScaleFormer, a novel architecture designed for multi-scale pansharpening that reframes generalization across image resolutions as generalization across sequence lengths.
94. Data-Centric Benchmark for Label Noise Estimation and Ranking in Remote Sensing Image Segmentation
Core Problem: High-quality pixel-level annotations are essential for semantic segmentation of remote sensing imagery, but they are expensive and often affected by noise, which significantly degrades the performance and robustness of modern segmentation models.
Key Innovation: Introduces a novel Data-Centric benchmark, a new publicly available dataset, and two techniques for identifying, quantifying, and ranking training samples according to their level of label noise in remote sensing semantic segmentation, leveraging model uncertainty, prediction consistency, and representation analysis.
95. Adapt Data to Model: Adaptive Transformation Optimization for Domain-shared Time Series Foundation Models
Core Problem: Large time series models (LTMs) struggle with the diversity and nonstationarity of real-world time series data, leading to poor trade-offs between forecasting accuracy and generalization across domains, often requiring costly fine-tuning.
Key Innovation: Proposes TATO, a data-centric framework that optimizes transformation pipelines (context slicing, scale normalization, outlier correction) to adapt a single frozen pre-trained LTM to diverse downstream time series domains, significantly improving forecasting performance and efficiency.
96. Retrodictive Forecasting: A Proof-of-Concept for Exploiting Temporal Asymmetry in Time Series Prediction
Core Problem: Conventional forward time series prediction may not fully exploit temporal asymmetry, potentially limiting forecasting accuracy in certain irreversible processes.
Key Innovation: Proposes 'retrodictive forecasting,' identifying the future that best explains the present via inverse MAP optimization over a CVAE, theoretically grounded in an information-theoretic arrow-of-time measure, demonstrating competitive or superior performance on irreversible time series, including climate reanalysis data.
97. RC-GeoCP: Geometric Consensus for Radar-Camera Collaborative Perception
Core Problem: The synergy between cameras and 4D radar for collaborative perception (CP) in multi-agent systems is underexplored, and existing methods struggle with misalignment due to depth ambiguity and spatial dispersion, especially in adverse weather where LiDAR-centric systems degrade.
Key Innovation: Introduces RC-GeoCP, the first framework for 4D radar and image fusion in CP, establishing a radar-anchored geometric consensus through Geometric Structure Rectification, Uncertainty-Aware Communication, and a Consensus-Driven Assembler, achieving state-of-the-art performance with reduced communication overhead.
98. TokenSplat: Token-aligned 3D Gaussian Splatting for Feed-forward Pose-free Reconstruction
Core Problem: Joint 3D Gaussian reconstruction and camera pose estimation from unposed multi-view images is challenging, particularly in feed-forward architectures without iterative refinement, leading to lower fidelity and unstable pose estimation.
Key Innovation: Presents TokenSplat, a feed-forward framework for joint 3D Gaussian reconstruction and camera pose estimation, introducing a Token-aligned Gaussian Prediction module, learnable camera tokens, and an Asymmetric Dual-Flow Decoder for higher fidelity and improved pose accuracy in pose-free settings.
99. Neural Latent Arbitrary Lagrangian-Eulerian Grids for Fluid-Solid Interaction
Core Problem: Effectively capturing highly nonlinear two-way Fluid-Solid Interaction (FSI) problems remains a significant challenge, with most deep learning methods limited to simplified one-way scenarios or struggling with dynamic, heterogeneous interactions due to a lack of cross-domain awareness.
Key Innovation: Introduces Fisale, a data-driven framework inspired by classical ALE and partitioned coupling methods, which explicitly models the coupling interface, leverages multiscale latent ALE grids for unified geometry-aware embeddings, and uses a partitioned coupling module for progressive modeling of nonlinear interdependencies. It handles complex two-way FSI problems and excels in reality-related scenarios.
100. Navigating Time's Possibilities: Plausible Counterfactual Explanations for Multivariate Time-Series Forecast through Genetic Algorithms
Core Problem: Understanding and modeling causality in complex and dynamic multivariate time series to uncover hidden causal relationships and identify potential interventions for desired outcomes, particularly in the context of counterfactual learning.
Key Innovation: Introduces a novel method for counterfactual learning in multivariate time series forecasting, integrating genetic algorithms with Granger causality tests to infer and validate causal dependencies, and using genetic algorithms with quantile regression to project future scenarios under hypothetical interventions, providing plausible counterfactual explanations.
101. Spectral Super-Resolution via Adversarial Unfolding and Data-Driven Spectrum Regularization: From Multispectral Satellite Data to NASA Hyperspectral Image
Core Problem: Multispectral satellite data (e.g., Sentinel-2) has limited spectral and non-unified spatial resolution, restricting its utility, while high-resolution hyperspectral data (e.g., AVIRIS-NG) has limited global coverage, posing a challenge for achieving global hyperspectral coverage.
Key Innovation: Proposes UALNet, a novel deep unfolding framework that achieves spectral super-resolution (12-to-186 bands) and unifies spatial resolution (to 5m) from Sentinel-2 data to NASA hyperspectral images. It uses a data-driven spectrum prior from PriorNet and integrates an adversarial term for guided reconstruction, outperforming existing methods in efficiency and accuracy.
102. Adaptive Augmentation-Aware Latent Learning for Robust LiDAR Semantic Segmentation
Core Problem: Adverse weather conditions significantly degrade LiDAR point cloud semantic segmentation performance by introducing large distribution shifts, and existing augmentation methods struggle to fully exploit augmentations due to the trade-off between minor and aggressive augmentations and the resulting semantic shift.
Key Innovation: A3Point, an adaptive augmentation-aware latent learning framework that effectively utilizes diverse augmentations while mitigating semantic shift. It consists of semantic confusion prior (SCP) latent learning and semantic shift region (SSR) localization, enabling adaptive optimization strategies for different disturbance levels.
103. GRAD-Former: Gated Robust Attention-based Differential Transformer for Change Detection
Core Problem: Existing deep learning approaches for change detection in remote sensing struggle to precisely delineate change regions, especially with very high-resolution (VHR) satellite images due to quadratic computational complexity and poor performance with limited training data.
Key Innovation: GRAD-Former, a novel framework with an Adaptive Feature Relevance and Refinement (AFRAR) module, uses gated mechanisms and differential attention to enhance contextual understanding and efficiently delineate change regions in VHR satellite images, outperforming state-of-the-art models across multiple datasets with fewer parameters.
104. Operator Learning Using Weak Supervision from Walk-on-Spheres
Core Problem: Training neural PDE solvers is bottlenecked by expensive data generation or unstable physics-informed neural networks (PINNs) that involve challenging optimization landscapes due to higher-order derivatives.
Key Innovation: Proposes WoS-NO (Walk-on-Spheres Neural Operator), a learning scheme that uses weak supervision from Monte Carlo walk-on-spheres to train neural operators. This amortizes the cost of Monte Carlo walks, resulting in a mesh-free, data-free physics-informed objective that generalizes to novel PDE parameters and domains with significant improvements in error, speed, and memory.
105. Continuous Exposure-Time Modeling for Realistic Atmospheric Turbulence Synthesis
Core Problem: Existing methods for synthesizing atmospheric turbulence effects oversimplify the relationship between blur and exposure-time, leading to unrealistic synthetic data and limited generalization for long-range imaging.
Key Innovation: Proposes a novel Exposure-Time-dependent MTF (ET-MTF) to model blur as a continuous function of exposure-time, derives a tilt-invariant PSF, and constructs ET-Turb, a large-scale synthetic turbulence dataset, resulting in more realistic restorations and superior generalization on real-world data.
106. One Operator to Rule Them All? On Boundary-Indexed Operator Families in Neural PDE Solvers
Core Problem: Neural PDE solvers implicitly learn boundary-indexed families of operators rather than a single boundary-agnostic operator, leading to fundamental limitations in generalization when boundary conditions vary.
Key Innovation: Formalizes this limitation by framing operator learning as conditional risk minimization, demonstrating that generalization in forcing terms or resolution does not imply generalization across boundary conditions, highlighting the need for explicit boundary-aware modeling in neural PDE solvers.
107. Towards OOD Generalization in Dynamic Graphs via Causal Invariant Learning
Core Problem: Existing dynamic graph neural networks (DyGNNs) struggle with out-of-distribution (OOD) shifts, failing to identify invariant patterns, capture intrinsic evolution rationale, and generalize across diverse OOD shifts with limited data.
Key Innovation: Proposes Dynamic graph Causal Invariant Learning (DyCIL) which uses a dynamic causal subgraph generator, a causal-aware spatio-temporal attention module, and an adaptive environment generator to exploit invariant spatio-temporal patterns from a causal view for robust OOD generalization.
108. PromptStereo: Zero-Shot Stereo Matching via Structure and Motion Prompts
Core Problem: While modern stereo matching methods leverage monocular depth foundation models for zero-shot generalization, the crucial iterative refinement stage remains underexplored, and conventional GRU-based architectures struggle to effectively exploit monocular depth priors as guidance.
Key Innovation: Proposes Prompt Recurrent Unit (PRU), a novel iterative refinement module based on monocular depth foundation model decoders, which integrates monocular structure and stereo motion cues as prompts to enrich latent representations with absolute stereo-scale information, achieving state-of-the-art zero-shot generalization.
109. Unifying Heterogeneous Multi-Modal Remote Sensing Detection Via Language-Pivoted Pretraining
Core Problem: Heterogeneous multi-modal remote sensing object detection suffers from unstable training and suboptimal generalization due to the entanglement of modality alignment and task-specific optimization in existing late alignment paradigms.
Key Innovation: Proposes BabelRS, a unified language-pivoted pretraining framework that decouples modality alignment from downstream task learning using Concept-Shared Instruction Aligning (CSIA) to align modalities to shared linguistic concepts and Layerwise Visual-Semantic Annealing (LVSA) for fine-grained semantic guidance, stabilizing training and outperforming state-of-the-art methods.
110. Accelerating PDE Surrogates via RL-Guided Mesh Optimization
Core Problem: Deep surrogate models for parametric PDEs require extensive data (thousands of fine-grid simulations), leading to high computational costs and limiting practical deployment.
Key Innovation: RLMesh, an end-to-end framework using reinforcement learning to adaptively allocate mesh grid points non-uniformly, focusing resolution in critical regions. This significantly reduces simulation queries while maintaining accuracy, making PDE surrogates more efficient for a wide range of problems, including potential geohazard models.
111. From Pixels to Patches: Pooling Strategies for Earth Embeddings
Core Problem: Default mean pooling for aggregating pixel-level geospatial foundation model embeddings into patch representations discards within-patch variability, leading to significant accuracy drops and reduced geographic generalization.
Key Innovation: Evaluation of 11 training-free and 2 parametric pooling methods on EuroSAT-Embed, showing that richer pooling schemes (like Generalized Mean Pooling (GeM) or Stats pooling) significantly reduce the geographic generalization gap and increase accuracy compared to mean pooling, especially for higher-dimensional embeddings, thus improving geospatial data processing.
112. 3D Field of Junctions: A Noise-Robust, Training-Free Structural Prior for Volumetric Inverse Problems
Core Problem: Many 3D imaging inverse problems, including those involving lidar point clouds, face high levels of measurement noise, and existing denoising methods may require training data or struggle to preserve sharp structures.
Key Innovation: A novel, fully volumetric 3D Field of Junctions (3D FoJ) representation that serves as a noise-robust, training-free structural prior for volumetric inverse problems, effectively denoising and enhancing sharp edge and corner structures in 3D data, including lidar point clouds, even under low signal-to-noise ratios.
113. GeoDiT: Point-Conditioned Diffusion Transformer for Satellite Image Synthesis
Core Problem: Existing controlled satellite image generative models often rely on time-consuming and semantically limited pixel-level maps for conditioning, hindering flexible and annotation-friendly control over the generation process.
Key Innovation: GeoDiT, a diffusion transformer for text-to-satellite image generation that uses a novel point-based conditioning framework, allowing for semantically rich and flexible control via spatial points and associated textual descriptions, and incorporating an adaptive local attention mechanism for improved performance.
114. Agentic Scientific Simulation: Execution-Grounded Model Construction and Reconstruction
Core Problem: Natural-language descriptions of physics-based simulation models are underspecified, leading to ambiguities that hinder correctness and reproducibility when using LLM agents for model construction.
Key Innovation: JutulGPT, an agentic scientific simulation framework built on a differentiable reservoir simulator, organizes model construction as an execution-grounded interpret-act-validate loop, detecting and resolving underspecified choices autonomously or via user queries, and demonstrating a methodology for auditing reproducibility.
115. PhysFormer: A Physics-Embedded Generative Model for Physically Self-Consistent Spectral Synthesis
Core Problem: Modeling high-dimensional complex systems governed by partial differential equations (PDEs) often lacks physical consistency and numerical stability, as existing methods rely on known physical fields or external loss functions.
Key Innovation: Introduces PhysFormer, a generative modeling framework that embeds physical processes directly within the network, learning key physical quantities from data in a low-dimensional latent space to ensure physical consistency and enhance spectral fidelity and inversion stability in complex systems, highly applicable to geohazard modeling.
116. Tiny-DroNeRF: Tiny Neural Radiance Fields aboard Federated Learning-enabled Nano-drones
Core Problem: Enabling complex vision-based tasks like dense 3D scene reconstruction on resource-constrained nano-drones, as top-performing Neural Radiance Fields (NeRF) models require significant memory and computation, typically delivered by high-end GPUs.
Key Innovation: Introduces Tiny-DroNeRF, a lightweight NeRF model optimized for ultra-low-power MCUs aboard nano-drones, achieving a 96% memory footprint reduction. It further empowers this with a collaborative federated learning scheme, distributing model training among multiple drones to increase overall reconstruction accuracy and overcome single-drone memory limitations.
117. TRAKNN: Efficient Trajectory Aware Spatiotemporal kNN for Rare Meteorological Trajectory Detection
Core Problem: Capturing the temporal evolution (trajectory) of atmospheric circulation patterns is essential for characterizing rare and impactful atmospheric behavior, but exhaustive similarity search on multi-decadal, continental-scale gridded datasets is computationally and memory intensive.
Key Innovation: Proposes TRAKNN, an unsupervised and data-agnostic framework for detecting geometrically rare short trajectories in spatio-temporal data using an exact kNN approach, leveraging a recurrence-based algorithm and efficient batch operations to enable exhaustive analysis on standard workstations.
118. DAWA: Dynamic Ambiguity-Wise Adaptation for Real-Time Domain Adaptive Semantic Segmentation
Core Problem: Existing Test-Time Domain Adaptation (TTDA) methods for semantic segmentation suffer from poor adaptation efficiency and continuous semantic ambiguities due to costly frame-wise optimization or unrealistic domain shift assumptions.
Key Innovation: DAWA, a real-time framework for TTDA semantic segmentation that adaptively detects domain shifts and dynamically adjusts learning strategies using a Dynamic Ambiguous Patch Mask (DAP Mask) and Dynamic Ambiguous Class Mix (DAC Mix), outperforming state-of-the-art methods while maintaining real-time inference speeds.
119. InvAD: Inversion-based Reconstruction-Free Anomaly Detection with Diffusion Models
Core Problem: Existing reconstruction-based anomaly detection methods using diffusion models suffer from a trade-off between fidelity and efficiency due to fine-grained noise-strength tuning and computationally expensive multi-step denoising.
Key Innovation: Proposes InvAD, an inversion-based, reconstruction-free anomaly detection approach that directly infers the final latent variable via DDIM inversion in a few steps, measuring deviation from a known prior for anomaly scoring, achieving state-of-the-art performance with approximately 2x inference-time speedup.
120. Flexible-weighted Chamfer Distance: Enhanced Objective Function for Point Cloud Completion
Core Problem: The standard Chamfer Distance (CD) in point cloud completion uses a symmetric weighting, leading to structural defects like point aggregation and incomplete spatial structures by equally penalizing local detail and global coverage.
Key Innovation: Flexible-weighted Chamfer Distance (FCD), an enhanced objective function that decouples CD into local precision and global completeness sub-objectives with an asymmetric weighting strategy, prioritizing global structural integrity to produce more uniform and structurally complete point clouds.
121. HGTS-Former: Hierarchical HyperGraph Transformer for Multivariate Time Series Analysis
Core Problem: Analyzing complex multivariate time series data remains challenging due to its high dimensionality, dynamic nature, and intricate interactions among variables.
Key Innovation: HGTS-Former, a novel hypergraph-based time series Transformer backbone network that uses hierarchical hypergraphs to aggregate temporal patterns within each channel and fine-grained relations between different variables, achieving state-of-the-art performance on multiple time series analysis tasks.
122. Semantic-Enhanced Time-Series Forecasting via Large Language Models
Core Problem: Existing LLM-based time series forecasting methods focus on token-level modal alignment, failing to bridge the intrinsic modality gap between linguistic knowledge structures and time series data patterns, thus limiting semantic representation. Additionally, Transformer-based LLMs are weak at modeling short-term anomalies.
Key Innovation: SE-LLM, a Semantic-Enhanced LLM that explores inherent periodicity and anomalous characteristics of time series to embed into the semantic space, enhancing token interpretability. It also includes a plugin module within self-attention to model both long-term and short-term dependencies, effectively adapting LLMs to time-series analysis.
123. Fifty Years of Object Detection and Recognition from Synthetic Aperture Radar Remote Sensing Imagery: The Road Forward
Core Problem: Unlocking the full potential of AI for Synthetic Aperture Radar (SAR) image understanding, given the historical challenges and the rapid evolution of AI technology, requires a clear roadmap and addressing critical bottlenecks in SAR Automatic Target Recognition (SAR ATR).
Key Innovation: Provides the first comprehensive review of SAR ATR over five decades, covering pivotal challenges, datasets, methods, and evaluation, and identifies promising future directions, emphasizing high-quality large-scale datasets, fair benchmarks, and open-source ecosystems to enable bidirectional empowerment between AI and SAR image understanding.
124. Towards Generalizable PDE Dynamics Forecasting via Physics-Guided Invariant Learning
Core Problem: Deep learning methods for forecasting PDE dynamics struggle with zero-shot out-of-distribution (OOD) generalization across capricious real-world physical environments and unseen PDE system parameters, as they often lack integration of fundamental physical invariance.
Key Innovation: Explicitly defined a two-fold PDE invariance principle (ingredient operators and their composition relationships remain invariant). Proposed iMOOE, a physics-guided invariant learning method with an Invariance-aligned Mixture Of Operator Expert architecture and a frequency-enriched invariant learning objective, achieving superior in-distribution performance and zero-shot OOD generalization for PDE dynamics forecasting.
125. UniMMAD: Unified Multi-Modal and Multi-Class Anomaly Detection via MoE-Driven Feature Decompression
Core Problem: Existing anomaly detection (AD) methods treat modality and class independently, leading to fragmented solutions, excessive memory overhead, and struggles with large variations across domains in multi-class reconstruction-based approaches, resulting in distorted normality boundaries and high false alarm rates.
Key Innovation: UniMMAD, a unified framework for multi-modal and multi-class anomaly detection, which employs a Mixture-of-Experts (MoE)-driven feature decompression mechanism for adaptive and disentangled reconstruction, achieving state-of-the-art performance across diverse AD datasets with improved efficiency.
126. Measuring the Intrinsic Dimension of Earth Representations
Core Problem: A lack of understanding regarding the information content and concentration within geographic Implicit Neural Representations (INRs) used for Earth observation, hindering their evaluation and design.
Key Innovation: Provides the first study of the intrinsic dimensionality of geographic INRs, demonstrating that this metric can quantify information content, correlate with downstream task performance, and capture spatial artifacts, offering an architecture-agnostic, label-free metric for unsupervised evaluation and model selection in Earth observation.
127. ReSAM: Refine, Requery, and Reinforce: Self-Prompting Point-Supervised Segmentation for Remote Sensing Images
Core Problem: Interactive segmentation models like SAM perform suboptimally on remote sensing imagery (RSI) due to severe domain shifts and the scarcity of dense annotations, limiting their applicability for scalable analysis.
Key Innovation: A point-supervised self-prompting framework (ReSAM) that adapts SAM to RSI using only sparse point annotations, employing a Refine-Requery-Reinforce loop to progressively enhance segmentation quality and domain robustness through self-guided prompt adaptation and Soft Semantic Alignment.
128. Deformation-Free Cross-Domain Image Registration via Position-Encoded Temporal Attention
Core Problem: Cross-domain image registration is challenging due to coupled geometric misalignment and domain-specific appearance shifts, often requiring explicit deformation field estimation.
Key Innovation: GPEReg-Net, a deformation-free framework that formalizes registration as a factorization problem (decomposing images into domain-invariant scene representation and global appearance statistic) and introduces a position-encoded cross-frame attention mechanism to exploit temporal coherence, achieving state-of-the-art performance on image registration benchmarks.
129. Astral: training physics-informed neural networks with error majorants
Core Problem: The primal approach to training Physics-informed Neural Networks (PiNNs) relies on residual minimization, which is an indirect measure of solution error, making it difficult to reliably estimate accuracy and stop optimization when desired accuracy is reached.
Key Innovation: Astral loss, a novel training method for PiNNs that uses error majorants to provide a direct upper bound on the solution error, leading to faster convergence, lower error, and, crucially, the ability to reliably estimate and spatially correlate with the true error, which is impossible with residual-based losses.
130. Learning sparsity-promoting regularizers for linear inverse problems
Core Problem: Selecting effective sparsity-promoting regularizers for linear inverse problems often relies on heuristic choices or fixed operators, limiting the ability to optimally leverage data-specific statistical properties and prior knowledge.
Key Innovation: Introduces a novel bilevel optimization framework to learn an an optimal synthesis operator (B) for sparsity-promoting regularization in linear inverse problems, leveraging data statistics and prior knowledge, and extending previous Tikhonov regularization efforts to non-differentiable norms and data-driven sparse regularization.
131. A Benchmark Dataset for Machine Learning Surrogates of Pore-Scale CO2-Water Interaction
Core Problem: Accurately capturing the complex, pore-scale interaction between CO2 and water in porous media is crucial for geoscience applications like carbon capture and storage, but high-fidelity simulations are computationally intensive, necessitating machine learning surrogates.
Key Innovation: Introduces a comprehensive, high-fidelity numerical simulation dataset of 624 2D samples (512x512, 100 time steps) capturing pore-scale CO2-water interaction under varying heterogeneity, providing high-resolution temporal and spatial information for benchmarking machine learning models in geoscience applications.
132. SamudrACE: Fast and Accurate Coupled Climate Modeling with 3D Ocean and Atmosphere Emulators
Core Problem: Traditional numerical global climate models are computationally intensive, limiting the speed and resolution of centuries-long simulations of the full Earth system.
Key Innovation: Presents SamudrACE, a coupled global climate model emulator using machine learning, which produces centuries-long simulations at high resolution with 3D ocean and atmosphere components, demonstrating high stability, low climate biases, and realistic variability in coupled climate phenomena.
133. AD-HKFCM: A Robust Nonlinear Spectral Variability-Aware Unmixing via Intra/Inter-Class Affinity Cohesion
Core Problem: Spectral variability and nonlinear mixing interactions significantly degrade the accuracy of spectral unmixing, especially in complex, heterogeneous environments.
Key Innovation: Proposes AD-HKFCM, a robust nonlinear spectral variability-aware unmixing model that integrates fuzzy clustering, a hybrid kernel function, support vector data description-derived hypersphere centers, and a physics-informed affinity distance metric to precisely infer "virtual pure endmembers" and achieve significantly lower abundance estimation errors.
134. Adaptive Edge-Guided Fusion Network for Remote Sensing Image Change Detection
Core Problem: Existing remote sensing image change detection (CD) networks often insufficiently utilize shallow-layer features and have limited extraction of edge information, despite deepening network depth.
Key Innovation: Proposes an adaptive edge-guided fusion CD network that uses an edge-guided branch with wavelet transformation to filter and fuse rich edge information, a cross-scale interaction module based on Transformers for multiscale feature fusion, and a hierarchical fusion module in the decoding stage, outperforming state-of-the-art methods on representative RS datasets.
135. Mamba-FCS: Joint Spatio-Frequency Feature Fusion, Change-Guided Attention, and SeK Inspired Loss for Enhanced Semantic Change Detection in Remote Sensing
Core Problem: Semantic Change Detection (SCD) requires models that integrate extensive spatial context, computational efficiency for large datasets, and sensitivity to class-imbalanced land-cover transitions, which traditional CNNs and Transformers struggle with due to limited receptive fields or quadratic complexity.
Key Innovation: Introduces Mamba-FCS, an SCD framework leveraging a Visual State Space Model backbone. It features a Joint Spatio-Frequency Fusion block to sharpen edges and mitigate illumination artifacts, a Change-Guided Attention (CGA) module to bridge Binary Change Detection and SCD tasks, and a novel SeK-inspired loss function to optimize for class imbalance, consistently outperforming state-of-the-art algorithms.
136. High Precision Altimetry Using Spaceborne Grazing Angle GNSS-Reflectometry Over Tonlé Sap Lake
Core Problem: Measuring surface height variations of inland water bodies, especially under vegetation, where traditional methods like SWOT and ICESat-2 show increased noise.
Key Innovation: Demonstrates the use of spaceborne grazing-angle GNSS-Reflectometry (GNSS-R) for high-precision altimetry, maintaining similar noise levels under vegetation as over open water, outperforming other satellite methods in vegetated areas, and showing utility for improving geoid representations over large lakes.
137. Correction: Remote sensing of mountain snow from space: status and recommendations
Core Problem: This item is a journal correction notice that updates previously published material rather than presenting new primary geohazard evidence.
Key Innovation: Improves record accuracy and reproducibility; it does not add new experimental, observational, or modeling results.
138. Application of distributed fiber-optic sensing in mining pressure and overburden monitoring in three-dimensional similarity simulation experiments
Core Problem: Traditional monitoring methods for mining pressure and overburden deformation lack continuous internal deformation data, hindering precise monitoring and early warning for safe coal extraction.
Key Innovation: Integrates PPP-BOTDA distributed fiber-optic sensing with a self-developed traction displacement device in 3D similarity simulation experiments to continuously monitor internal deformation. It proposes a method to identify roof pressure and quantify overburden failure height using Brillouin frequency shift, demonstrating high sensitivity to small internal rock layer deformations.
139. Harnessing machine learning and multi-source data fusion for urban fire risk assessment: predictive analysis of spatial heterogeneity
Core Problem: The probability and risk of urban fires are highly uncertain due to multiple influencing factors and complex urban layouts, making effective risk assessment challenging.
Key Innovation: Developed an Urban Area Fire Probability Prediction (UAFPP) model using machine learning and multi-source data fusion to quantitatively analyze and predict urban fire risk, demonstrating robust predictive performance and providing actionable insights for mitigation.
140. Multidisciplinary analysis of the park fire: advancing mega-fire management
Core Problem: California's increasing vulnerability to mega-fires, driven by climate change and complex management challenges, leads to significant deficiencies in inter-agency collaboration, resource allocation, and community preparedness.
Key Innovation: Conducted a multidisciplinary analysis of the 2024 Park Fire using CAL FIRE reports, NASA satellite imagery, and governmental policies to evaluate progression, impacts, and response, providing recommendations for enhancing real-time predictive modeling, communication, and community engagement to improve resilience.
141. Impact of loading platen rotation and initial specimen tilting on drained triaxial compression of granular soils: DEM insights
Core Problem: Variations in triaxial testing apparatus configurations (e.g., loading platen rotation) and procedures (e.g., initial specimen tilting) can introduce measurement deviations in determining the mechanical properties of granular soils.
Key Innovation: Used the Discrete Element Method (DEM) to investigate how non-rotatable loading platens and initial specimen tilting influence the mechanical responses of granular soils. Found that non-rotatable platens significantly increased deviatoric stress and volumetric dilation in dense specimens, leading to 'X'-type shear bands, highlighting the impact of fabric anisotropy.
142. Quantifying particle morphology via lightweight object detection and generative data augmentation
Core Problem: Traditional methods for quantifying particle morphology in granular soils are inefficient for field-scale characterization, and existing deep learning approaches are computationally intensive, limiting mobile deployment.
Key Innovation: Proposed YOLOv8-PMQ, a lightweight object detection model for end-to-end prediction of particle shape and size distribution without image segmentation, by introducing a morphology parameter regression branch. Utilized an OASIS generative adversarial network (GAN) for data augmentation and developed an offline smartphone application for rapid on-site analysis.
143. Experimental and Numerical Investigation of Segment Response to Seismic Wave Propagation for Advanced Geological Detection in Shield Tunnels
Core Problem: Lack of universally applicable geological detection methods for shield tunneling, especially in challenging conditions like fracture zones and groundwater, leading to 'blind' advancement and hindering seismic detection due to absence of exposed rock.
Key Innovation: Investigated seismic wave propagation characteristics using segment-excited/received waves via full-scale experiments and numerical simulations. Verified the feasibility of segment-based seismic wave excitation/reception, providing theoretical and technical support for advanced geological detection in shield tunnels, and suggested practical improvements.
144. Experimental and Mechanistic Investigation on Time-Delayed Initiation of Hydraulic Fractures in Shale Under Constant Pressure Injection
Core Problem: Understanding the mechanisms and developing predictive models for time-delayed initiation of hydraulic fractures, especially in shale, which is critical for reducing breakdown pressure and suppressing seismic activity.
Key Innovation: Conducted true triaxial fracturing experiments on Lushan shale, using AE monitoring and microscopic observations. Showed constant pressure injection reduces breakdown pressure and disperses hydraulic energy, transitioning failure from instantaneous mechanical instability to fluid-rock coupling controlled. Developed a breakdown lifetime prediction model considering fluid infiltration, stress corrosion, and confining pressure.
145. Mechanism and Prediction of the Kaiser Effect in Rocks Under Historical Deep Thermo-mechanical Coupling Environments
Core Problem: Accurate in-situ stress measurement is essential for underground safety, but high-temperature and high-stress environments in deep mining challenge traditional methods, requiring a better understanding of the Kaiser effect under these conditions.
Key Innovation: Conducted rock compression tests with real-time reconstruction of deep thermo-mechanical coupling history using granite, combined with PFC2D simulations. Systematically analyzed the Kaiser effect's characteristics, evolution, and mechanisms. Developed a prestress prediction model to correct errors for in-situ stress testing in high-temperature environments, achieving high accuracy.
146. Discrete Element Study on Crack Development Laws Around Holes Considering the Bi-Modularity of Rocks
Core Problem: Understanding the law of crack development around holes in rock, considering the distinct bimodularity characteristics of rock materials, which directly affect stress concentration and hole failure.
Key Innovation: Developed a discrete element method (DEM) with a calibrated rock crack bond (RCB) contact model to investigate crack development under bimodularity effects. Showed that tensile cracks accumulate damage, while shear cracks induce and guide propagation. Analyzed the influence of hole inclination and interaction on crack patterns and stress bridges.
147. Mechanical Response and Damage Evolution of Marble Under Triaxial Constraint and Cyclic Loading–Unloading
Core Problem: Understanding the mechanical behavior, damage evolution, and failure modes of surrounding rock (marble) under triaxial constraint and cyclic loading–unloading conditions is crucial for evaluating deep rock mass stability during construction.
Key Innovation: Quantified cyclic damage using multiple methods, established an exponential damage evolution model accounting for confining pressure and three-stage damage. Revealed stress memory effect, non-linear evolution of elastic modulus and Poisson's ratio, and identified shear failure as dominant, providing a predictive framework for deep rock mass stability.
148. Examining the Case for Accuracy and Precision When Determining the Hoek–Brown Parameters
Core Problem: The Hoek–Brown failure criterion, despite its widespread use, suffers from unvalidated empirical foundations, non-uniqueness in parameter determination (e.g., m_b), and inherent qualitative uncertainty in GSI, leading to potentially inaccurate rock engineering designs.
Key Innovation: Critically analyzed the limitations of the Hoek–Brown criterion, highlighting the lack of field-scale validation, the non-uniqueness of m_b, and the inconsistency of pursuing precision for m_i given qualitative GSI. Emphasized the need for transparency, renewed field-scale validation, and caution against AI amplifying unvalidated assumptions in rock engineering.
149. Mixed-Mode I-II Fracture Mechanical Behaviors of Frozen Sandstone with Different Initial Saturation Degrees
Core Problem: The complex mixed-mode I-II fracture failure of rock in low-temperature environments and with varying initial saturation degrees poses significant challenges for rock mass engineering in cold regions.
Key Innovation: Investigated the influence of pore phase composition and crack dip angles on mixed-mode I-II fracture mechanics of frozen sandstone. Revealed two-stage variations in peak force, fracture toughness, and fracture energy with initial saturation, and phased nonlinear responses of fracture toughness to crack dip angle, attributing these to pore ice content and dynamic stress field changes.
150. Correction: Experimental Investigation and a Method of Calculating Bearing Capacity for Jointed Rock Masses Under Excavated and Unexcavated Conditions
Core Problem: This item is a journal correction notice that updates previously published material rather than presenting new primary geohazard evidence.
Key Innovation: Improves record accuracy and reproducibility; it does not add new experimental, observational, or modeling results.
151. Structural Response of Reinforced Concrete Frame Structures Subjected to Tunnelling Induced Differential Foundation Settlements
Core Problem: Lack of thorough understanding of the impact of tunnelling-induced differential settlements on the structural integrity and load redistribution in existing RC frame structures, especially considering soil-structure interactions and various structural geometries.
Key Innovation: Utilizes advanced nonlinear modeling techniques (PLAXIS 3D, ABAQUS) to investigate the structural response of RC frames to tunnelling-induced settlements, clarifying the role of structural configuration and providing insights for safe urban construction.
152. Numerical Modeling of Semi Ductile Failure in Intact and Jointed Rock Using Bonded Block Modeling Methods
Core Problem: The limitation of existing Bonded Block Modeling (BBM) methods in accurately replicating semi-ductile mechanical behavior in intact rock and predicting the strength of jointed rock specimens.
Key Innovation: Evaluation of three different BBM configurations for Carrara marble, demonstrating that current BBM approaches struggle to fully match ductile rock deformation and are not predictive for jointed rock strength, suggesting areas for improvement in rock mechanics modeling.
153. Influence of axial loads on the behavior of lightweight precast shear wall joints under cyclic loading
Core Problem: Optimizing the seismic performance of lightweight precast shear wall systems, specifically understanding the influence of axial loads on the behavior of their joints under cyclic loading to enhance strength, ductility, and energy dissipation.
Key Innovation: Experimental cyclic loading tests on lightweight Dual Connection Shear Wall systems, demonstrating the significant influence of axial load on seismic performance and identifying an optimal axial load ratio for enhanced ductility and energy dissipation, while utilizing sustainable lightweight concrete.
154. Downscaling GRACE(−FO) with Mass-Conserving XGBoost approach reveals High-Resolution patterns and drivers of Hydrometeorological-Induced mass changes in high Mountain Asia
Core Problem: The coarse resolution of GRACE(−FO) satellite data impedes the quantification of glacier mass changes at fine spatial scales in High Mountain Asia, which is critical for understanding climate impacts and regional water security.
Key Innovation: Proposed an XGBoost downscaling method integrated with mass conservation correction to derive high-resolution glacier mass changes. The method significantly reduced RMSE and outperformed existing Mascon solutions. Identified a pronounced atmospheric oscillation over the Caspian–Black Sea region driving interannual moisture transport and a distinct 6–7 year interannual mass oscillation in the Tien Shan–Pamir region.
155. An improved Hydrology-Informed attention LSTM(HIA-LSTM) model for runoff simulation with seasonal snowmelt
Core Problem: Accurate runoff simulation in alpine basins of the Tibetan Plateau remains challenging due to complex cryospheric processes and strong non-linear interactions among precipitation, snowmelt, and glacier melt, limiting conventional 'black-box' deep learning models.
Key Innovation: Proposed a Hydrology-Informed Attention LSTM (HIA-LSTM) model that embeds physical inductive biases (specialized attention heads for quickflow, slowflow, snowmelt; temporal masking, logarithmic decay, temperature-based gating) into the neural architecture. Achieved superior performance (average KGE 0.888, NSE 0.892) over standard LSTMs, particularly in melt-driven watersheds, and enhanced interpretability through attention weight analysis.
156. A dynamic double phase field model for simulating the frictional shear fractures and mixed-mode fractures in rock materials
Core Problem: Replicating complex fracturing patterns, including tensile and shear cracks with frictional contact, in rock engineering under dynamic loads is challenging with existing models.
Key Innovation: Proposes a dynamic double phase field model using two phase fields to simulate tensile and shear cracks, establishing a local coordinate system to judge contact states and decompose stress tensors, effectively simulating mixed-mode fractures and frictional contact in rock materials.
157. Deep Seafloor Ambient Seismic Noise Monitoring for Oceanic and Atmospheric Pressure Changes
Core Problem: Application of seismic noise interferometry in ocean environments has been limited by the scarcity of long-term ocean-bottom seismic recordings, and the relationship between deep seafloor seismic responses and atmospheric/oceanic pressure changes was unclear.
Key Innovation: Demonstrated that deep-seafloor seismic responses are sufficiently sensitive to capture atmospheric activity and pressure variations, highlighting the potential of passive seismic sensing for monitoring oceanographic and atmospheric processes and associated hazards.
158. Knowledge-guided generative surrogate modeling for high-dimensional design optimization under scarce data
Core Problem: Surrogate models for mechanical design and manufacturing optimization are often limited by data scarcity, and few techniques can systematically integrate valuable domain knowledge with limited data.
Key Innovation: Proposed RBF-Gen, a knowledge-guided surrogate modeling framework that combines scarce data with domain knowledge by constructing a radial basis function (RBF) space with more centers than training samples and leveraging the null space via a generator network. This encodes structural relationships and distributional priors, significantly outperforming standard RBF surrogates in data-scarce settings.
159. Efficient Image Super-Resolution with Multi-Scale Spatial Adaptive Attention Networks
Core Problem: Existing image super-resolution (SR) methods often face a dilemma between achieving high reconstruction fidelity and maintaining low model complexity, making them less efficient for practical applications requiring both quality and speed.
Key Innovation: Introduces the Multi-scale Spatial Adaptive Attention Network (MSAAN), a lightweight SR network that uses a novel Multi-scale Spatial Adaptive Attention Module (MSAA) to jointly model fine-grained local details and long-range contextual dependencies, achieving superior or competitive performance with significantly lower parameters and computational costs.
160. Predicting Local Climate Zones using Urban Morphometrics and Satellite Imagery
Core Problem: The efficacy of using urban morphometrics, alone or fused with satellite imagery, to predict Local Climate Zones (LCZs) needs evaluation, as the relationship between LCZs and measurable urban form aspects is tenuous.
Key Innovation: Evaluation of morphometric-based and fusion-based (morphometrics + satellite imagery) LCZ prediction schemes, using 321 2D morphometric attributes, demonstrating that a broader range of urban form properties are relevant for distinguishing LCZ types, though fusion benefits are modest and inconsistent.
161. Physics-Consistent Diffusion for Efficient Fluid Super-Resolution via Multiscale Residual Correction
Core Problem: Existing image super-resolution (SR) and generic diffusion models perform poorly in fluid SR, being sampling-intensive, ignoring physical constraints, and often yielding spectral mismatch and spurious divergence.
Key Innovation: Proposes ReMD (Residual-Multigrid Diffusion), a physics-consistent diffusion framework that performs a multigrid residual correction at each reverse step, coupling data consistency with lightweight physics cues and correcting residuals across scales using a multi-wavelet basis, leading to accelerated convergence, improved accuracy, and spectral fidelity with fewer sampling steps.
162. Zero-Shot and Supervised Bird Image Segmentation Using Foundation Models: A Dual-Pipeline Approach with Grounding DINO~1.5, YOLOv11, and SAM~2.1
Core Problem: Bird image segmentation is challenging due to pose diversity, plumage, and lighting, and traditional methods often require extensive retraining for new species or domains.
Key Innovation: A dual-pipeline framework for bird image segmentation leveraging foundation models (SAM 2.1, Grounding DINO 1.5, YOLOv11) that achieves high accuracy in both zero-shot and supervised modes with minimal fine-tuning, outperforming prior baselines and demonstrating domain adaptation via prompt-based methods.
163. Weakly Supervised Video Anomaly Detection with Anomaly-Connected Components and Intention Reasoning
Core Problem: Weakly supervised video anomaly detection (WS-VAD) struggles to learn anomaly semantics effectively due to the absence of dense frame-level annotations.
Key Innovation: LAS-VAD, a novel framework integrating an anomaly-connected component mechanism (to group semantically similar frames) and an intention awareness mechanism (to distinguish similar normal/abnormal behaviors), further enhanced by incorporating anomaly attribute information.
164. Exploring Spatiotemporal Feature Propagation for Video-Level Compressive Spectral Reconstruction: Dataset, Model and Benchmark
Core Problem: Existing spectral compressive imaging reconstruction methods are image-based, leading to uncertainty in reconstructing missing information and lacking temporal consistency for dynamic scenes.
Key Innovation: Proposes a video-level spectral reconstruction approach (PG-SVRT) leveraging spatiotemporal features and temporal continuity across frames, along with a new dynamic hyperspectral image dataset (DynaSpec), achieving superior reconstruction quality and temporal consistency.
165. Exploring 3D Dataset Pruning
Core Problem: Dataset pruning for 3D data is largely unexplored, and the long-tail class distribution in 3D datasets makes optimization under conventional metrics (OA, mAcc) conflicting and pruning challenging.
Key Innovation: Proposes a 3D dataset pruning method that formulates pruning as approximating expected risk, addresses coverage error and prior-mismatch bias through representation-aware subset selection with per-class retention quotas, and uses prior-invariant teacher supervision, improving both OA and mAcc.
166. PPC-MT: Parallel Point Cloud Completion with Mamba-Transformer Hybrid Architecture
Core Problem: Existing point cloud completion methods struggle to achieve both high-quality reconstruction and computational efficiency, particularly for unordered point clouds.
Key Innovation: Proposes PPC-MT, a novel parallel framework for point cloud completion that leverages a hybrid Mamba-Transformer architecture and a parallel completion strategy guided by Principal Component Analysis (PCA) to transform unordered point clouds into ordered subsets for parallel reconstruction, significantly enhancing uniformity, detail fidelity, and computational efficiency.
167. pySpatial: Generating 3D Visual Programs for Zero-Shot Spatial Reasoning
Core Problem: Multi-modal Large Language Models (MLLMs) struggle with tasks requiring spatial understanding of the 3D world, limiting their ability to reason explicitly over structured spatial representations.
Key Innovation: Introduces pySpatial, a visual programming framework that equips MLLMs with the ability to interface with spatial tools (e.g., 3D reconstruction, camera-pose recovery) via Python code generation, enabling zero-shot spatial reasoning in 3D environments.
168. Foundation Models in Remote Sensing: Evolving from Unimodality to Multimodality
Core Problem: The exponential growth and diversity of remote sensing (RS) data necessitate advanced data modeling and understanding capabilities for effective management and interpretation.
Key Innovation: A comprehensive technical survey on foundation models in RS, offering a new perspective on their evolution from unimodality to multimodality, and providing a tutorial-like guide for researchers to apply these models across various RS applications.
169. RaUF: Learning the Spatial Uncertainty Field of Radar
Core Problem: Millimeter-wave radar suffers from low spatial fidelity, azimuth ambiguity, and clutter, leading to ill-posed geometric inference and unreliable spatial perception, especially in adverse weather conditions.
Key Innovation: Proposes RaUF, a spatial uncertainty field learning framework that models radar measurements using physically grounded anisotropic properties and an anisotropic probabilistic model for fine-grained uncertainty. It also introduces a Bidirectional Domain Attention mechanism to suppress spurious reflections, leading to more reliable spatial detections.
170. LLaDA-o: An Effective and Length-Adaptive Omni Diffusion Model
Core Problem: The need for an effective and length-adaptive omni diffusion model for unified multimodal understanding and generation that addresses computational redundancy and supports flexible-length decoding in multimodal settings.
Key Innovation: LLaDA-o, an omni diffusion model built on a Mixture of Diffusion (MoD) framework that decouples discrete masked diffusion for text and continuous diffusion for visual generation, while coupling them through a shared attention backbone. It also introduces a data-centric length adaptation strategy for flexible-length decoding.
171. Flow Matching-enabled Test-Time Refinement for Unsupervised Cardiac MR Registration
Core Problem: Diffusion-based unsupervised image registration for cardiac cine MR is limited by expensive multi-step inference, hindering its practical application despite strong registration capabilities.
Key Innovation: FlowReg, a flow-matching framework in displacement field space that achieves strong registration in as few as two steps and supports further refinement. It uses warmup-reflow training and an Initial Guess strategy to improve efficiency and performance without needing a pre-trained model.
172. Unified Vision-Language Modeling via Concept Space Alignment
Core Problem: The need for a unified vision-language embedding space and model that supports a wide range of text and speech languages, enabling robust multilingual and multimodal understanding and generation capabilities.
Key Innovation: V-SONAR, a vision-language embedding space extended from the text-only SONAR space, constructed via a post-hoc alignment pipeline. This enables competitive text-to-video retrieval and state-of-the-art video captioning. Further, V-LCM extends the Large Concept Model (LCM) with vision-language instruction tuning, achieving state-of-the-art performance across 61 languages for image/video captioning and Q&A.
173. VP-Hype: A Hybrid Mamba-Transformer Framework with Visual-Textual Prompting for Hyperspectral Image Classification
Core Problem: Accurate hyperspectral imagery (HSI) classification is frustrated by the tension between high-dimensional spectral data and extreme scarcity of labeled training samples, compounded by the quadratic complexity of standard Transformers.
Key Innovation: Introduces VP-Hype, a hybrid Mamba-Transformer framework that unifies linear-time efficiency of State-Space Models with relational modeling of Transformers. Addresses label scarcity by integrating dual-modal Visual and Textual Prompts, achieving new state-of-the-art performance in low-data regimes for HSI classification.
174. Velocity Model Building and Editing with Guided Denoising Diffusion Implicit Models
Core Problem: Velocity-model building in seismic imaging is a challenging inverse problem due to limited data coverage, nonlinearity, and the need to integrate heterogeneous information, leading to less realistic velocity structures with classical inversion.
Key Innovation: A unified framework combining learned diffusion priors with structurally preconditioned inverse formulations (DDIM inversion and guided sampling) for velocity-model editing and full building, recovering sharper and more realistic velocity structures than classical inversion.
175. Perspective-Equivariant Fine-tuning for Multispectral Demosaicing without Ground Truth
Core Problem: Reconstructing full-resolution spectral images from snapshot mosaiced measurements (multispectral demosaicing) is challenging, with classical methods producing blurry results and supervised learning requiring costly ground truth data.
Key Innovation: PEFD (Perspective-Equivariant Fine-tuning for Demosaicing), a framework that learns multispectral demosaicing from mosaiced measurements alone by exploiting the projective geometry of camera systems and adapting pretrained foundation models, achieving fine detail recovery and spectral fidelity comparable to supervised methods without ground truth.
176. UTICA: Multi-Objective Self-Distllation Foundation Model Pretraining for Time Series Classification
Core Problem: The underexplored potential of non-contrastive self-supervised methods for time series foundation models, despite their significant advances in computer vision.
Key Innovation: UTICA, a multi-objective self-distillation foundation model for time series classification that adapts DINOv2-style self-distillation. It learns representations capturing both temporal invariance via augmented crops and fine-grained local structure via patch masking, achieving state-of-the-art classification performance on UCR and UEA benchmarks.
177. Fed-GAME: Personalized Federated Learning with Graph Attention Mixture-of-Experts For Time-Series Forecasting
Core Problem: Existing federated learning (FL) methods for distributed time-series forecasting on graphs rely on static topologies and struggle with client heterogeneity, limiting personalized aggregation.
Key Innovation: Fed-GAME, a personalized federated learning framework that models personalized aggregation as message passing over a learnable dynamic implicit graph. It employs a decoupled parameter difference-based update protocol and a novel Graph Attention Mixture-of-Experts (GAME) aggregator for fine-grained personalization, outperforming state-of-the-art baselines on real-world electric vehicle charging datasets.
178. Invariant-Stratified Propagation for Expressive Graph Neural Networks
Core Problem: Graph Neural Networks (GNNs) are fundamentally limited in expressivity by the 1-dimensional Weisfeiler-Leman (1-WL) test and fail to capture structural heterogeneity, leading to uniform information aggregation.
Key Innovation: Introduces Invariant-Stratified Propagation (ISP), a framework comprising ISP-WL and ISPGNN, which stratifies nodes according to graph invariants to process them in hierarchical strata, enhancing expressivity beyond 1-WL and improving performance across graph tasks.
179. UETrack: A Unified and Efficient Framework for Single Object Tracking
Core Problem: Most existing single object tracking methods are limited to RGB inputs, struggle in multi-modal scenarios, and are too heavy and slow for resource-constrained deployment.
Key Innovation: Proposes UETrack, an efficient, unified framework for multi-modal single object tracking, utilizing a Token-Pooling-based Mixture-of-Experts and a Target-aware Adaptive Distillation strategy to achieve a superior speed-accuracy trade-off across various benchmarks and modalities.
180. Tackling multiphysics problems via finite element-guided physics-informed operator learning
Core Problem: Efficiently solving multiphysics problems with coupled partial differential equations (PDEs) on arbitrary domains, especially for discretization-independent prediction without relying on labeled simulation data, remains a challenge.
Key Innovation: A finite element-guided physics-informed operator learning framework (Folax) that learns a mapping from input parameter space to solution space using a weighted residual formulation, enabling discretization-independent prediction beyond training resolution and exploring various neural operator backbones for complex geometries.
181. WildCross: A Cross-Modal Large Scale Benchmark for Place Recognition and Metric Depth Estimation in Natural Environments
Core Problem: Existing robotics datasets are predominantly captured in structured urban environments, making them inadequate for addressing the challenges of complex, unstructured natural settings for tasks like place recognition and metric depth estimation.
Key Innovation: Proposes WildCross, a cross-modal benchmark dataset comprising over 476K sequential RGB frames with semi-dense depth, surface normal annotations, 6DoF poses, and synchronized dense lidar submaps, specifically designed for place recognition and metric depth estimation in large-scale natural environments.
182. Benchmarking Semantic Segmentation Models via Appearance and Geometry Attribute Editing
Core Problem: The need to thoroughly test and improve the robustness of semantic segmentation models in varied and complex scenes, especially concerning variations in appearance and geometry attributes, which existing evaluation paradigms do not fully address.
Key Innovation: Gen4Seg, an automatic data generation pipeline that uses diffusion models to edit visual attributes (appearance and geometry) of real images with precise control, enabling stress-testing and benchmarking of semantic segmentation models and serving as a data augmentation tool to improve robustness.
183. Training-Free Spatio-temporal Decoupled Reasoning Video Segmentation with Adaptive Object Memory
Core Problem: The computational cost and resource demands of fine-tuning MLLMs for Reasoning Video Object Segmentation (ReasonVOS), and the temporal instability caused by coupled spatio-temporal information processing in existing methods.
Key Innovation: SDAM, a training-free spatio-temporal decoupled reasoning video segmentation framework that uses only pre-trained models, incorporating an Adaptive Object Memory module for key object selection and memorization, and spatio-temporal decoupling for stable cross-frame propagation, achieving excellent results on benchmark datasets.
184. DeLo: Dual Decomposed Low-Rank Experts Collaboration for Continual Missing Modality Learning
Core Problem: Adapting Large Multimodal Models (LMMs) to real-world scenarios faces challenges in continual learning from sequential data streams and handling frequent missing modalities, with existing prompt tuning or naive LoRA methods suffering from cross-task/modality interference.
Key Innovation: Proposes DeLo, a novel dual-decomposed low-rank expert architecture for Continual Missing Modality Learning (CMML) that resolves modality interference through decomposed LoRA experts, a Cross-Modal Guided Routing strategy, and a Task-Key Memory, significantly outperforming state-of-the-art approaches.
185. FreeGNN: Continual Source-Free Graph Neural Network Adaptation for Renewable Energy Forecasting
Core Problem: Accurate forecasting of renewable energy generation is crucial, but traditional supervised models require labeled data from target sites, which is often unavailable, and models need to continually adapt to non-stationary environmental conditions without source data or target labels.
Key Innovation: Proposes FreeGNN, a Continual Source-Free Graph Domain Adaptation framework that integrates a spatio-temporal GNN with a teacher-student strategy, memory replay, graph-based regularization, and a drift-aware weighting scheme, enabling continuous adaptation to non-stationary conditions for robust renewable energy forecasting on unseen sites.
186. A Practical Guide to Streaming Continual Learning
Core Problem: Real-world problems often require machine learning agents to both rapidly adapt to changes in data streams (concept drifts, as in Streaming Machine Learning) and retain past knowledge when learning new tasks (as in Continual Learning), but existing approaches alone struggle to achieve both simultaneously.
Key Innovation: Introduces Streaming Continual Learning (SCL) as an emerging paradigm that unifies Continual Learning (CL) and Streaming Machine Learning (SML) challenges, proposing it as a solution to foster hybrid approaches that can achieve both rapid adaptation and knowledge retention, thereby connecting the two communities and addressing limitations of individual approaches.
187. Streaming Continual Learning for Unified Adaptive Intelligence in Dynamic Environments
Core Problem: Developing effective predictive models is challenging in dynamic environments that continuously produce data and constantly change, requiring adaptation to non-stationary data streams without forgetting previous knowledge.
Key Innovation: Streaming Continual Learning (SCL), a unified setting that harnesses the benefits of both Continual Learning and Streaming Machine Learning to enable models to quickly adapt to non-stationary data streams while retaining previous knowledge, extending techniques from both fields.
188. Meta-Learning Hyperparameters for Parameter Efficient Fine-Tuning
Core Problem: Parameter-Efficient Fine-Tuning (PEFT) methods for adapting large foundation models to domain-specific applications like remote sensing are highly sensitive to fixed hyperparameters, hindering performance, especially on tail classes.
Key Innovation: Proposes MetaPEFT, a method that incorporates adaptive scalers to dynamically adjust key PEFT factors (module insertion, layer selection, module-wise learning rates) during fine-tuning, achieving state-of-the-art performance in cross-spectral adaptation for remote sensing with improved tail-class accuracy.
189. DGNet: Discrete Green Networks for Data-Efficient Learning of Spatiotemporal PDEs
Core Problem: Neural PDE solvers typically require large amounts of expensive high-fidelity training data and struggle with data efficiency and generalization to unseen source terms because they don't explicitly encode the strong structural inductive biases of PDE dynamics.
Key Innovation: Proposes DGNet, a discrete Green network for data-efficient learning of spatiotemporal PDEs, which transforms Green's function into a graph-based discrete formulation and embeds the superposition principle into a hybrid physics-neural architecture, achieving state-of-the-art accuracy with limited data and robust zero-shot generalization.
190. Near-Field Focusing Operators for Planar Multi-Static Microwave Imaging Using Back-Projection in the Spatial Domain
Core Problem: Traditional back-projection methods for planar multi-static microwave imaging, especially in near-field scenarios, often fail to accurately reconstruct images due to neglecting magnitude correction factors and suffering from imaging artifacts.
Key Innovation: Derives an improved formalism for microwave image creation using back-projection in the spatial domain, including analytically derived integral expressions for focusing operators that provide magnitude correction factors. This procedure is shown to be superior to traditional methods, particularly in near-field, and effectively suppresses imaging artifacts through low-pass filtering.
191. Consistent initialization of mixed-dimensional multiphysics models for fractured reservoirs under geomechanical constraints and field measurements
Core Problem: Accurately initializing mixed-dimensional multiphysics models for fractured reservoirs, particularly the in-situ stress state and fracture aperture, is challenging due to the mismatch between idealized computational models and real-world deformed configurations, leading to large deviations in flow predictions.
Key Innovation: A discrete fracture model with constitutive laws expressed relative to the unknown equilibrium state, paired with a fixed-point initialization strategy that consistently reconstructs the reference configuration based on geomechanical constraints and field measurements, improving accuracy for fractured reservoir modeling.
192. physfusion: A Transformer-based Dual-Stream Radar and Vision Fusion Framework for Open Water Surface Object Detection
Core Problem: Detecting water-surface targets for USVs is challenging due to wave clutter, specular reflections, weak appearance cues, and the sparse/intermittent nature of maritime radar point clouds with heavy-tailed reflectivity variations, making conventional fusion designs ineffective.
Key Innovation: Proposes PhysFusion, a physics-informed radar-image detection framework. It integrates a Physics-Informed Radar Encoder (PIR Encoder) for robust feature learning under clutter, a Radar-guided Interactive Fusion Module (RIFM) for query-level radar-image fusion, and a Temporal Query Aggregation module (TQA) for temporally consistent representations, achieving improved detection performance on water-surface targets.
193. ORGAN: Object-Centric Representation Learning using Cycle Consistent Generative Adversarial Networks
Core Problem: Extracting information from images by segmenting them into objects and representing each object in a low-dimensional latent space, especially for challenging real-world datasets with many objects and low visual contrast.
Key Innovation: ORGAN, a novel object-centric representation learning approach based on cycle-consistent Generative Adversarial Networks, which performs similarly to state-of-the-art on synthetic data and uniquely handles challenging real-world datasets, scales well, and creates expressive latent space representations.
194. SimRecon: SimReady Compositional Scene Reconstruction from Real Videos
Core Problem: Conventional compositional scene reconstruction approaches from real-world videos primarily emphasize visual appearance, leading to limited generalization, visual infidelity of generated assets, and physical implausibility when assembling object-centric representations for simulation.
Key Innovation: Proposes SimRecon, a 'Perception-Generation-Simulation' pipeline for cluttered scene reconstruction, which includes Active Viewpoint Optimization (to acquire optimal projected images for single-object completion) and a Scene Graph Synthesizer (to guide physically plausible construction in 3D simulators), addressing visual fidelity and physical plausibility.
195. OnlineX: Unified Online 3D Reconstruction and Understanding with Active-to-Stable State Evolution
Core Problem: Existing 3D reconstruction methods, including advanced ones like 3D Gaussian Splatting, primarily follow an offline paradigm, lacking the capacity for continuous online reconstruction and suffering from cumulative drift due to the conflicting demands of capturing high-frequency local geometry and preserving long-term global structure.
Key Innovation: Introduces OnlineX, a feed-forward framework for unified online 3D visual appearance and language field reconstruction from streaming images, which addresses cumulative drift through a decoupled active-to-stable state evolution paradigm and incorporates an implicit Gaussian fusion module for enhanced quality and real-time inference.
196. Is Bigger Always Better? Efficiency Analysis in Resource-Constrained Small Object Detection
Core Problem: The assumption that larger models trained on more data consistently outperform smaller ones, leading to inefficient model selection in resource-constrained Earth observation (EO) for small object detection.
Key Innovation: A systematic efficiency analysis demonstrating that smaller, high-resolution models (YOLO11N) can achieve both higher efficiency and absolute performance for small object detection in EO, with input resolution being the dominant resource allocation lever and additional data yielding negligible returns at low resolution.
197. Tipping the Balance: Impact of Class Imbalance Correction on the Performance of Clinical Risk Prediction Models
Core Problem: The impact of class-imbalance correction techniques on the probabilistic calibration of ML-based clinical risk prediction models, especially in settings with rare outcomes, is not sufficiently understood.
Key Innovation: A comprehensive evaluation across ten clinical datasets and multiple ML models showing that commonly used class-imbalance correction strategies (SMOTE, RUS, ROS) did not systematically improve discrimination and generally degraded calibration performance, suggesting caution in their application for probabilistic risk prediction.
198. Time-Aware Latent Space Bayesian Optimization
Core Problem: Most Latent-space Bayesian Optimization (LSBO) methods assume a fixed objective, failing to account for temporal drift in real design campaigns, which can affect both the surrogate model and the latent search space geometry.
Key Innovation: Proposes Time-Aware Latent-space Bayesian Optimization (TALBO), which incorporates time into both the surrogate and the learned generative representation via a GP-prior variational autoencoder, yielding a latent space aligned as objectives evolve and outperforming baselines on drifting multi-property molecular design tasks.
199. Non-Rectangular Average-Reward Robust MDPs: Non-Rectangular Average-Reward Robust MDPs:Optimal Policies and Their Transient Values
Core Problem: Designing optimal policies for robust Markov decision processes (MDPs) under the average-reward criterion, especially when ambiguity sets are non-rectangular and couple transition probabilities, and ensuring good finite-time performance (transient values) beyond just average-reward optimality.
Key Innovation: Shows that history-dependent policies achieving sublinear expected regret are robust-optimal, establishes a minimax representation for the robust value without rectangularity, and constructs an epoch-based policy that combines an optimal stationary policy with online learning to achieve a constant-order transient value.
200. Thoth: Mid-Training Bridges LLMs to Time Series Understanding
Core Problem: Large Language Models (LLMs) struggle to understand and reason about time series data, limiting their effectiveness in decision-making scenarios that depend on temporal dynamics.
Key Innovation: Thoth, the first family of mid-trained LLMs with general-purpose time series understanding capabilities, achieved through a pivotal intermediate mid-training stage using 'Book-of-Thoth', a high-quality, time-series-centric corpus for task- and domain-agnostic alignment between time series and natural language.
201. SEAnet: A Deep Learning Architecture for Data Series Similarity Search
Core Problem: Existing SAX-based indexes for data series similarity search perform poorly with high-frequency, noisy, or weakly correlated data, limiting effective analysis of massive data series collections.
Key Innovation: Proposes SEAnet, a deep learning architecture for Deep Embedding Approximation (DEA) that incorporates a Sum of Squares preservation property and novel sampling strategies, enabling high-quality data series summarization and similarity search, with potential applications in geohazard monitoring.
202. Reconstructing Content via Collaborative Attention to Improve Multimodal Embedding Quality
Core Problem: The causal attention and next-token prediction paradigm of MLLMs, while effective for generation, does not explicitly encourage the formation of globally compact representations, limiting their effectiveness as multimodal embedding backbones.
Key Innovation: Proposes CoCoA, a Content reconstruction pre-training paradigm based on Collaborative Attention, which restructures attention flow and introduces an EOS-based reconstruction task to encourage multimodal models to generate compact and informative representations, significantly improving embedding quality for diverse data relevant to geohazards.
203. FedHB: Hierarchical Bayesian Federated Learning
Core Problem: Developing a robust and privacy-preserving federated learning framework that can model client-specific data while benefiting from a global model, and ensuring theoretical guarantees.
Key Innovation: Proposing FedHB, a hierarchical Bayesian federated learning approach whose variational inference leads to a distributed algorithm compatible with FL, subsuming existing algorithms, and offering convergence and generalization error analyses.
204. Topological Inductive Bias fosters Multiple Instance Learning in Data-Scarce Scenarios
Core Problem: The effectiveness of Multiple Instance Learning (MIL) significantly decreases in data-scarce scenarios, hindering its application in fields like rare disease classification.
Key Innovation: Introducing Topology Guided MIL (TG-MIL), which incorporates topological inductive biases into the data representation space to enhance MIL classifier performance and generalizability, especially under scarce-data regimes, demonstrating improvements across various datasets.
205. FiLo: Zero-Shot Anomaly Detection by Fine-Grained Description and High-Quality Localization
Core Problem: Existing zero-shot anomaly detection methods struggle with precisely matching diverse anomaly types and accurately localizing anomalies of various sizes and scales due to generic descriptions and single-patch feature comparisons.
Key Innovation: FiLo, a novel zero-shot anomaly detection (ZSAD) method using adaptively learned Fine-Grained Descriptions (FG-Des) via LLMs and position-enhanced High-Quality Localization (HQ-Loc) with a Multi-scale Multi-shape Cross-modal Interaction (MMCI) module, significantly improving detection and localization performance.
206. MV-Adapter: Enhancing Underwater Instance Segmentation via Adaptive Channel Attention
Core Problem: Existing underwater instance segmentation models struggle to effectively adapt to feature variations across different channels in complex underwater environments, leading to suboptimal performance due to issues like light attenuation, color distortion, and complex backgrounds.
Key Innovation: The MarineVision Adapter (MV-Adapter) module, which introduces an adaptive channel attention mechanism to dynamically adjust feature weights based on underwater image characteristics, thereby enhancing the model's ability to handle light attenuation, color shifts, and complex backgrounds, leading to improved segmentation performance.
207. LLaVE: Large Language and Vision Embedding Models with Hardness-Weighted Contrastive Learning
Core Problem: Existing LMM-based embedding models trained with standard InfoNCE loss exhibit high overlap in similarity distribution between positive and negative pairs, making it challenging to effectively distinguish hard negative pairs in multimodal tasks.
Key Innovation: LLaVE, a simple yet effective framework that dynamically improves the embedding model's representation learning for negative pairs based on their discriminative difficulty, achieving state-of-the-art performance in multimodal benchmarks and demonstrating strong scalability and zero-shot generalization.
208. SemHiTok: A Unified Image Tokenizer via Semantic-Guided Hierarchical Codebook for Multimodal Understanding and Generation
Core Problem: Previous unified image tokenizers struggle to achieve a good trade-off between capturing high-level semantic features for multimodal understanding and retaining low-level pixel features for generation due to differing feature priorities in joint training.
Key Innovation: SemHiTok, a unified image tokenizer with a novel semantic-guided hierarchical codebook that decouples semantic and pixel features, enabling it to capture pixel features while retaining high-level semantic comprehension, leading to leading performance in image reconstruction and multimodal understanding.
209. VideoFusion: A Spatio-Temporal Collaborative Network for Multi-modal Video Fusion
Core Problem: Lack of large-scale multi-sensor video datasets and difficulty in jointly modeling spatial and temporal dependencies for multi-modal video fusion, leading to temporal inconsistency in existing image-oriented fusion methods.
Key Innovation: Constructs M3SVD, a benchmark dataset for multi-sensor video fusion, and proposes VideoFusion, a spatio-temporal collaborative network that uses a differential reinforcement module, a complete modality-guided fusion strategy, and a bi-temporal co-attention mechanism to generate spatio-temporally coherent videos, outperforming image-oriented fusion paradigms.
210. GradPCA: Leveraging NTK Alignment for Reliable Out-of-Distribution Detection
Core Problem: Existing Out-of-Distribution (OOD) detection methods often lack consistent performance across different benchmarks, and a robust theoretical framework for designing principled spectral OOD detectors is needed.
Key Innovation: GradPCA, an OOD detection method that applies Principal Component Analysis (PCA) to gradient class-means, leveraging the low-rank structure of neural network gradients induced by Neural Tangent Kernel (NTK) alignment, achieving more consistent performance and providing a theoretical framework for spectral OOD detection.
211. Point-MoE: Large-Scale Multi-Dataset Training with Mixture-of-Experts for 3D Semantic Segmentation
Core Problem: Scaling 3D point cloud understanding with large-scale, multi-dataset joint training for semantic segmentation is challenging due to heterogeneous data (sensors, scenes, densities, semantic biases), which degrades standard models when naively mixed.
Key Innovation: Introduces Point-MoE, a Mixture-of-Experts design that expands model capacity through sparsely activated expert MLPs and a lightweight router, allowing tokens to select specialized experts without requiring dataset supervision. It outperforms prior methods on diverse indoor/outdoor datasets and in zero-shot settings.
212. Rapid training of Hamiltonian graph networks using random features
Core Problem: Training Hamiltonian Graph Networks (HGN) for modeling complex N-body dynamics is slow due to iterative, gradient-descent-based optimization algorithms.
Key Innovation: Replaces iterative optimization with random feature-based parameter construction, achieving 150-600x faster training for HGNs with comparable accuracy, while retaining essential physical invariances in diverse simulations like N-body mass-spring and molecular dynamics.
213. Causal Structure Learning in Hawkes Processes with Complex Latent Confounder Networks
Core Problem: Uncovering causal structures and identifying latent subprocesses in complex temporal event systems modeled by Hawkes processes, especially when only partially observed.
Key Innovation: Proposes a two-phase iterative algorithm that leverages the representation of continuous-time event sequences as discrete-time causal models to infer causal relationships and uncover new latent subprocesses, with theoretical guarantees for identifiability. This method has strong potential for modeling geohazard event sequences like earthquakes or landslides.
214. Massively Multimodal Foundation Models: A Framework for Capturing Interactions with Specialized Mixture-of-Experts
Core Problem: Capturing complex, time-varying interactions, especially delayed cross-modal effects, across a large number of heterogeneous input streams (modalities) is challenging for existing Mixture-of-Experts (MoE) architectures, leading to suboptimal expert specialization and reduced accuracy.
Key Innovation: A framework that explicitly quantifies temporal dependencies between modality pairs and uses an interaction-aware router to guide MoE routing, enabling experts to learn generalizable interaction-processing skills and achieve substantial performance gains in massively multimodal settings.
215. Data-to-Energy Stochastic Dynamics
Core Problem: Existing algorithms for the Schr"odinger bridge problem require samples from both marginal distributions, limiting their applicability when one or both distributions are only known by their unnormalised densities (data-free scenario).
Key Innovation: The first general method for modeling Schr"odinger bridges when one or both distributions are given by unnormalised densities, relying on a generalization of the iterative proportional fitting (IPF) procedure to the data-free case, demonstrating efficacy in learning transports between multimodal distributions and data-free image-to-image translation.
216. Bayesian Influence Functions for Hessian-Free Data Attribution
Core Problem: Classical influence functions are difficult to apply to deep neural networks due to non-invertible Hessians and high-dimensional parameter spaces, limiting their utility for data attribution and understanding model behavior.
Key Innovation: The local Bayesian influence function (BIF), an extension that replaces Hessian inversion with loss landscape statistics estimated via stochastic-gradient MCMC sampling, providing a Hessian-free approach that captures higher-order interactions and scales efficiently to large neural networks for data attribution.
217. LVTINO: LAtent Video consisTency INverse sOlver for High Definition Video Restoration
Core Problem: Extending state-of-the-art image-based latent diffusion models (LDMs) to high-definition video restoration is challenging due to the need for fine spatial detail and temporal consistency, with naive frame-by-frame application leading to inconsistent reconstructions.
Key Innovation: Proposes LVTINO, the first zero-shot inverse solver for high-definition video restoration that leverages Video Consistency Models (VCMs) as priors, achieving state-of-the-art reconstruction quality with strong measurement consistency and smooth temporal transitions across frames.
218. ResCP: Reservoir Conformal Prediction for Time Series Forecasting
Core Problem: Existing conformal prediction methods for time series are complex, fail with small sample sizes, and require expensive retraining when the underlying data distribution changes.
Key Innovation: Proposes Reservoir Conformal Prediction (ResCP), a novel training-free method leveraging reservoir computing to dynamically reweight conformity scores, accounting for local temporal dynamics without compromising computational scalability, and achieving asymptotic conditional coverage.
219. LinearSR: Unlocking Linear Attention for Stable and Efficient Image Super-Resolution
Core Problem: Generative models for Image Super-Resolution (SR) are computationally bottlenecked by self-attention's quadratic complexity, and linear attention, while efficient, has faced stability and quality issues.
Key Innovation: LinearSR is a holistic framework that systematically overcomes challenges in applying linear attention to photorealistic SR, achieving state-of-the-art perceptual quality and exceptional efficiency through novel strategies like "knee point"-based Early-Stopping Guided Fine-tuning (ESGF) and an SNR-based Mixture of Experts (MoE).
220. Incomplete Multi-Label Image Recognition by Co-learning Semantic-Aware Features and Label Recovery
Core Problem: Multi-label image recognition with incomplete labels faces fundamental challenges in learning robust semantic-aware features and accurately recovering missing labels.
Key Innovation: CSL (Co-learning framework for Semantic-aware features and Label recovery) addresses both challenges in a unified paradigm, developing a semantic-related feature learning module, a semantic-guided feature enhancement module, and a collaborative learning framework that dynamically enhances feature discriminability and adaptively infers missing labels.
221. UniFlow: A Unified Pixel Flow Tokenizer for Visual Understanding and Generation
Core Problem: Existing unified visual tokenizers face a significant performance trade-off between high-level semantic abstraction for visual understanding and low-level pixel reconstruction for visual generation.
Key Innovation: UniFlow is a generic and unified pixel flow tokenizer that flexibly adapts any visual encoder with a concise reconstruction decoder, using layer-wise adaptive self-distillation and a lightweight patch-wise pixel flow decoder to achieve a win-win outcome in both visual understanding and generation tasks.
222. There is No VAE: End-to-End Pixel-Space Generative Modeling via Self-Supervised Pre-training
Core Problem: Pixel-space generative models (diffusion and consistency models) are often more difficult to train and generally underperform compared to their latent-space counterparts, leaving a persistent performance and efficiency gap.
Key Innovation: A novel two-stage training framework that uses self-supervised pre-training of encoders to capture meaningful semantics and align sampling trajectories, followed by end-to-end fine-tuning with a randomly initialized decoder, achieving state-of-the-art performance for pixel-space diffusion and consistency models without relying on VAEs.
223. Concerto: Joint 2D-3D Self-Supervised Learning Emerges Spatial Representations
Core Problem: Learning coherent and informative spatial features for 3D scene perception from multisensory data remains a challenge, with existing self-supervised models often lacking superior fine-grained geometric and semantic consistency.
Key Innovation: Concerto, a minimalist self-supervised learning framework that combines 3D intra-modal self-distillation with 2D-3D cross-modal joint embedding. It learns more coherent and informative spatial features, outperforming standalone SOTA 2D and 3D models and setting new SOTA results across multiple scene understanding benchmarks.
224. Parameterized Prompt for Incremental Object Detection
Core Problem: Existing prompt-pool-based approaches for incremental learning are unsuitable for incremental object detection (IOD) because they assume disjoint class sets and overlook object co-occurrence, leading to confusion and catastrophic forgetting.
Key Innovation: P$^2$IOD (Parameterized Prompts for Incremental Object Detection), an approach that leverages neural networks as parameterized prompts to adaptively consolidate knowledge across tasks. It employs a parameterized prompts fusion strategy to constrain structure updates, preventing confusion and catastrophic forgetting, and achieves state-of-the-art performance in IOD.
225. UrbanFM: Scaling Urban Spatio-Temporal Foundation Models
Core Problem: Urban computing remains fragmented due to 'scenario-specific' models that are overfitted to specific regions or tasks, hindering their generalizability for dynamic spatio-temporal data streams in urban systems.
Key Innovation: UrbanFM, a minimalist self-attention architecture, is proposed as a spatio-temporal foundation model for urban systems, supported by WorldST (a billion-scale corpus), MiniST (a novel split mechanism), and EvalST (a large-scale benchmark), demonstrating remarkable zero-shot generalization across unseen cities and tasks.
226. On weight and variance uncertainty in neural networks for regression tasks
Core Problem: Existing Bayesian neural network approaches for regression often treat variance as fixed or deterministic, limiting the model's ability to fully capture and express predictive uncertainty, which can hinder generalization and reliability.
Key Innovation: Extends the weight uncertainty framework in neural networks by explicitly incorporating and modeling variance uncertainty, demonstrating that considering a full posterior distribution over the variance significantly enhances the predictive performance and generalization capability of Bayesian neural networks for regression tasks.
227. DoFlow: Flow-based Generative Models for Interventional and Counterfactual Forecasting on Time Series
Core Problem: Time-series forecasting increasingly demands not only accurate observational predictions but also causal forecasting under interventional and counterfactual queries in multivariate systems, along with principled anomaly detection.
Key Innovation: DoFlow, a flow-based generative model defined over a causal Directed Acyclic Graph (DAG), delivers coherent observational, interventional, and counterfactual predictions for time series, and provides explicit likelihoods for effective anomaly detection.
228. A Neural Network-Based Real-time Casing Collar Recognition System for Downhole Instruments
Core Problem: Autonomous and real-time casing collar recognition in downhole environments is challenging due to signal corruption from magnetic interference and stringent computational/power budgets.
Key Innovation: Collar Recognition Nets (CRNs), a family of lightweight 1-D convolutional neural networks optimized for efficiency and accuracy, demonstrating robust, autonomous, and real-time collar recognition on an embedded system with high throughput and low latency for downhole instruments.
229. Direct measurement of shear stress and shear strain in soil: getting more out of simple shear testing
Core Problem: Traditional simple shear testing relies on boundary measurements, leading to inaccuracies in determining small strains and stiffness degradation over the full range of relevant shear strains.
Key Innovation: Implementation of local, in-soil measurement tools for shear stress and strain, allowing for more accurate determination of stiffness degradation curves and maximum shear stiffness from a single simple shear test.
230. Classification of Seabed Types by OBIA With Recursive Feature Elimination Corrected by Expert Interpretation
Core Problem: Automated machine learning methods for seabed type classification often lack full consistency with reality and struggle with validation, while expert analysis is accurate but time-consuming and difficult to scale for large areas.
Key Innovation: Presents a data processing scheme that combines object-based image analysis (OBIA) with recursive feature elimination (RFE) and CatBoost, then refines the results using hybrid classes and expert interpretation, minimizing expert time and improving classification accuracy for seabed types.
231. Brownian Distance Covariance Prototype Refinement for Few-Shot Multitemporal Coastal Wetland Hyperspectral Image Classification
Core Problem: Accurate and long-term land cover monitoring of coastal wetlands is challenged by limited annotations in newly collected hyperspectral images and domain shifts from historical data, impairing model transferability.
Key Innovation: Proposes BDCPR-FSL, a novel cross-domain few-shot learning framework that uses a spectral-aware Transformer, Brownian distance covariance for nonlinear dependencies, a prototype refinement mechanism, and domain alignment loss to achieve superior classification performance for multitemporal coastal wetland hyperspectral images.
232. Hi-RSMamba: Hierarchical Mamba for Remote Sensing Image Restoration Under Adverse Weather
Core Problem: Remote sensing images are degraded by adverse weather, impairing quality and downstream task performance, with existing CNNs and Transformers having limitations in handling long-range dependencies and computational efficiency, respectively.
Key Innovation: Proposing Hi-RSMamba, a hierarchical state-space model that enhances contextual modeling by integrating multiscale representations (global and local state-space models) and uses a gated feedforward network for adaptive multiscale feature broadcast, achieving high-quality image restoration.
233. Getting the Most Out of the Image-Level Labels: (Un)Supervised Learning for Extracting Soil Parameters From Hyperspectral Images
Core Problem: Building effective (un)supervised AI models for estimating soil parameters from hyperspectral images is challenging due to the high cost of acquiring detailed ground truth, often leaving only coarse-grained, image-level labels.
Key Innovation: Proposing a comprehensive framework for estimating soil parameters from HSIs, including a spectrally- and spatially informed algorithm to generate superpixel-level pseudolabels from image-level labels, outperforming state-of-the-art supervised models and improving generalizability.
234. A Self-Supervised Conditional Diffusion Network With Multihead Wavelet Attention for Remote Sensing Image Denoising
Core Problem: Hyperspectral remote sensing images (HSIs) are degraded by multiple noise sources, limiting subsequent interpretation, and existing denoising methods struggle with reliance on paired clean data and balancing global structure preservation with local detail recovery.
Key Innovation: Proposing DM-WAN, a self-supervised conditional diffusion network with multihead wavelet attention, which incorporates a detail feature prompting strategy and a time-dependent feature modulation mechanism, and uses a multihead wavelet attention mechanism for collaborative multiscale feature fusion and structure-preserving reconstruction without requiring paired ground-truth supervision.
235. Toward More Adaptive Fusion: Sample- and Region-Aware Dynamic Fusion Network for Full- and Missing-Modalities Land Cover Classification
Core Problem: Existing methods for land cover classification using optical and SAR images struggle with sufficiently fusing multimodal features and reconstructing missing-modality features, often using fixed fusion strategies that lack adaptability to distinct sample pairs and variations in generation quality across regions.
Key Innovation: Proposing the Sample- and Region-Aware Dynamic Fusion Network (SRDFNet) which includes a multiexpert adaptive fusion module for sample-level adaptive feature fusion and a confidence-based adaptive feature generation module to refine generated features at the region level, significantly improving LCC performance under full- and missing-modality scenarios.
236. Hierarchically Optimized Forest Coverage Modeling With Multisource Remote Sensing Data for High-Precision Understory Terrain Mapping
Core Problem: The uncertainty in the accuracy of understory terrain severely restricts its applications, and precise terrain information is crucial but difficult to obtain, especially under dense forest cover.
Key Innovation: Proposes an innovative data fusion method integrating multisource remote sensing data (LiDAR, Sentinel, DEMs, forest parameters) with a forest-coverage-based hierarchical optimization strategy to construct a machine learning model for high-precision understory terrain estimation, improving accuracy by 51% compared to existing models.
237. CA-ConvLSTM: A Cross-Attention-Based Deep Learning Framework for Joint Prediction of Sea Surface Temperature and Sea Level Anomaly
Core Problem: Effectively capturing the complex dynamic coupling between sea surface temperature (SST) and sea level anomaly (SLA), along with their spatiotemporal dependencies and mutual influences, remains a significant challenge for accurate prediction.
Key Innovation: Proposes CA-ConvLSTM, a unified deep learning framework that jointly predicts SST and SLA by integrating a cross-attention mechanism with a ConvLSTM architecture. The cross-attention module enables dynamic feature interaction, allowing simultaneous learning from both variables through a unified training process, significantly improving prediction accuracy.
238. A New 3-D Positioning Method Using Zero-Doppler Image Matching for High-Squint Small SAR Satellite
Core Problem: Traditional SAR radargrammetry for 3-D positioning faces challenges with target matching due to varying viewing geometries and reliance on external topographic data, particularly with submeter-resolution SAR imagery.
Key Innovation: Proposes a novel 3-D positioning technique using only SAR images, without external data, by converting zero-Doppler image patches using a reference ellipsoid model for robust target matching and constructing a pixel-wise look-up table for precise extraction of Doppler centroid frequencies and slant ranges.
239. Assessing the impact of Earth Observation data-driven calibration of the melting coefficient on the LISFLOOD snow module
Core Problem: Improving the accuracy of snowmelt simulation within the LISFLOOD hydrological model, specifically by refining the snowmelt coefficient, which is typically calibrated against discharge data and may not adequately represent snow dynamics.
Key Innovation: Integration of Sentinel-2 and MODIS data to create an improved reference snow cover fraction (SCF) for calibrating a spatially distributed snowmelt coefficient in LISFLOOD, leading to improved bias and RMSE in snow cover representation (up to 8%) and noticeable divergences in discharge within smaller upstream catchments, highlighting the potential of EO data for targeted model component calibration.
240. A comparative study of time integration schemes for groundwater flow and transport using the differential quadrature method of lines
Core Problem: Accurate and efficient numerical modeling of groundwater flow and contaminant transport is essential, but traditional fixed-step schemes encounter stability limitations at high spatial resolutions.
Key Innovation: Presented a comprehensive numerical analysis of various time integration schemes combined with the Differential Quadrature Method (DQM) for groundwater flow and transport problems. Demonstrated that DQM with adaptive solvers maintains robustness and accuracy, yielding lower error margins than standard FDM with fewer grid nodes.
241. Field investigation of ammonia alkali soda residue-based foamed concrete as a subgrade filler
Core Problem: There is a need for high-value utilization of solid waste like ammonia alkali soda residue (AR) in construction, and the field performance of ammonia alkali soda residue-based foamed concrete (A-FC) as a subgrade filler has not been thoroughly investigated.
Key Innovation: Conducted a comprehensive field investigation of A-FC as a subgrade filler for an expressway, measuring key parameters like wet density, compressive strength, CBR, and settlement. Demonstrated its excellent road performance, meeting design requirements, and its potential for large-scale application in soft soil areas due to low weight and high strength.
242. Effect of Mechanical Anchorage Configuration on the Failure Behavior of Inclined Short Pile Foundation Joints
Core Problem: Ensuring adequate pull-out resistance and understanding failure mechanisms of inclined short pile foundations for transmission line towers in complex mountainous terrains, particularly concerning the optimal mechanical anchorage configuration.
Key Innovation: A comparative study of anchoring plate and hooked bar configurations for inclined short pile foundations, revealing that anchoring plates strengthen the joint and increase overall uplift bearing capacity, providing a reference for design in mountainous areas.
243. Enhancing transferability of foliar trait retrieval models: A comparative analysis of transfer learning strategies and domain shift characterization
Core Problem: The transferability of spectroscopic foliar trait retrieval models across diverse datasets is severely limited by domain shifts, and there is a lack of systematic comparison of transfer learning strategies and clear characterization of domain shift mechanisms.
Key Innovation: This study evaluated four categories of transfer learning strategies for retrieving five foliar traits across various ecological scenarios and quantitatively decomposed domain shift into its components, identifying concept shift as the primary factor limiting transferability, thereby providing insights for selecting tailored transfer strategies to advance robust remote sensing applications.
244. Incorporating learned geospatial embeddings to deep image prior for inpainting cloud areas in remotely sensed images
Core Problem: Cloud removal in optical remote sensing typically relies on extensive data and supervised training, limiting its applicability in data-scarce scenarios, and Deep Image Prior (DIP) alone often suffers from over-smoothing in large cloud-covered regions.
Key Innovation: This study combined ready-to-use Satellite Embedding data with DIP to enhance cloud removal and sub-cloud information reconstruction, demonstrating that reduced-dimensional embeddings and derived range/shape priors significantly improve spectral consistency and structure reconstruction performance in cloud-contaminated regions, thereby reducing data preparation burden and improving processing efficiency.
245. A FEM inspired unified algorithm for the drained and undrained elementary DEM tests of saturated sand under generalized loading path
Core Problem: Existing Discrete Element Method (DEM) studies often simulate undrained element tests under constant volume assumptions, limiting their application in simulating saturated sand behavior under complex drained and undrained loading paths.
Key Innovation: Proposes a unified algorithm combining FEM and DEM to simulate drained and undrained elementary tests of saturated soil under complex loading paths, describing macroscopic behavior via Biot's equation and using DEM for effective stress response, verified against laboratory tests.
246. Disentangling Wind‐ and Buoyancy‐Driven Changes in Pacific Barotropic Circulation and Regional Sea Level During 1960–2014
Core Problem: The relative roles of wind stress and buoyancy forcing in shaping Pacific circulation and regional sea level changes over the past 6 decades remain unclear.
Key Innovation: Used large-ensemble simulations to disentangle the contributions of wind and buoyancy fluxes, finding that wind stress accounts for 81% of barotropic circulation changes and 54% of regional sea-level trend, while buoyancy forcing contributes 19% and 46% respectively, highlighting the dominant yet regionally modulated role of wind stress.
247. StaTS: Spectral Trajectory Schedule Learning for Adaptive Time Series Forecasting with Frequency Guided Denoiser
Core Problem: Existing diffusion models for probabilistic time series forecasting use fixed noise schedules that hinder invertibility and structural recovery, and often lack modeling of schedule-induced spectral degradation.
Key Innovation: Introduced StaTS, a diffusion model that learns a data-adaptive noise schedule with spectral regularization (STS) and employs a Frequency Guided Denoiser (FGD) to improve structural preservation and heterogeneous restoration across diffusion steps, achieving consistent gains in forecasting performance.
248. Reinforcement Learning for Control with Probabilistic Stability Guarantee: A Finite-Sample Approach
Core Problem: A critical gap exists in reinforcement learning for control systems regarding providing rigorous stability guarantees, especially when learning from finite sampled data in a model-free framework.
Key Innovation: Developed a novel RL approach that provides probabilistic stability guarantees for control systems using finite data, leveraging Lyapunov's method to derive a probabilistic stability theorem and a policy gradient theorem, and introduced the L-REINFORCE algorithm, demonstrating its effectiveness in ensuring stability.
249. Property-Driven Evaluation of GNN Expressiveness at Scale: Datasets, Framework, and Study
Core Problem: The expressiveness of Graph Neural Networks (GNNs) in capturing fundamental graph properties remains an open challenge, hindering the development of trustworthy AI.
Key Innovation: Developed a property-driven evaluation methodology for GNNs, including a configurable graph dataset generator (GraphRandom, GraphPerturb) with 336 new datasets covering 16 graph properties, and a framework to assess generalizability, sensitivity, and robustness. Conducted a comprehensive study on global pooling methods, revealing distinct trade-offs.
250. BiJEPA: Bi-directional Joint Embedding Predictive Architecture for Symmetric Representation Learning
Core Problem: Standard Joint Embedding Predictive Architecture (JEPA) models rely on uni-directional prediction, potentially neglecting informative signals from inverse relationships, and symmetric prediction faces inherent instability (representation explosion).
Key Innovation: Proposed BiJEPA, a Bi-Directional Joint Embedding Predictive Architecture that enforces cycle-consistent predictability between data segments and introduces a critical norm regularization mechanism to address instability. It achieves stable convergence, captures semantic structure, and learns robust temporal and spatial representations.
251. Automated Quality Check of Sensor Data Annotations
Core Problem: Training AI algorithms for safety-critical applications like automated driving requires large amounts of high-quality, annotated sensor data, but manual quality assurance is labor-intensive, time-consuming, and prone to errors.
Key Innovation: Presents an automatic method and open-source tool for assuring the quality of multi-sensor training data, designed to detect nine common errors in railway vehicle datasets, significantly reducing manual workload and accelerating AI system development.
252. TinyVLM: Zero-Shot Object Detection on Microcontrollers via Vision-Language Distillation with Matryoshka Embeddings
Core Problem: Zero-shot object detection requires large vision-language models (VLMs) that exceed the memory constraints of microcontrollers (MCUs), preventing their deployment on resource-constrained edge devices.
Key Innovation: TinyVLM, the first framework for zero-shot object detection on MCUs with less than 1MB memory, featuring a decoupled architecture, Matryoshka distillation for nested embeddings, and quantized embedding storage.
253. Latent Replay Detection: Memory-Efficient Continual Object Detection on Microcontrollers via Task-Adaptive Compression
Core Problem: Existing continual learning methods for object detection require storing raw images, exceeding the memory budgets of microcontrollers (MCUs), preventing models from learning new object categories after deployment on edge devices.
Key Innovation: Latent Replay Detection (LRD), a framework for memory-efficient continual object detection on MCUs, featuring task-adaptive compression with FiLM conditioning, spatial-diverse exemplar selection, and an MCU-deployable system that stores latent representations instead of raw images.
254. AdaFocus: Knowing When and Where to Look for Adaptive Visual Reasoning
Core Problem: Existing training-free Multimodal Large Language Model (MLLM) methods for visual reasoning suffer from perceptual redundancy and a drift between semantic intent and spatial attention, leading to computational expense and inaccurate localization.
Key Innovation: AdaFocus, a novel training-free framework with a two-stage pipeline (confidence-based module for 'when to crop' and semantic-guided localization for 'where to crop'), achieving substantial performance gains and 4.0x speedup in adaptive visual reasoning.
255. Summer-22B: A Systematic Approach to Dataset Engineering and Training at Scale for Video Foundation Model
Core Problem: The engineering challenges and design decisions involved in systematically training a large-scale video foundation model from scratch, including dataset curation and scaling.
Key Innovation: A systematic approach to dataset engineering (metadata-driven curation, multi-stage filtering) and training (µP parameterization, hypersphere-constrained optimization) for the Summer-22B video foundation model, along with the Lavender Data system for dataset management.
256. Infinite Self-Attention
Core Problem: The quadratic cost of softmax attention limits Transformer scalability in high-resolution vision tasks.
Key Innovation: Infinite Self-Attention (InfSA), a spectral reformulation of self-attention as a diffusion process on a content-adaptive token graph, and Linear-InfSA, a linear-time variant that approximates the principal eigenvector, enabling stable training and inference at very high resolutions with significant efficiency gains.
257. Task-Driven Subspace Decomposition for Knowledge Sharing and Isolation in LoRA-based Continual Learning
Core Problem: Existing LoRA-based Continual Learning (CL) methods for sequential task adaptation overlook task-shared directions, suppressing knowledge transfer, and fail to capture truly effective task-specific directions due to inactive 'null bases' of old tasks.
Key Innovation: Proposed Low-rank Decomposition and Adaptation (LoDA) which performs a task-driven decomposition to build general and truly task-specific LoRA subspaces by solving energy-based objectives, decoupling directions for knowledge sharing and isolation, and using a Gradient-Aligned Optimization (GAO) approach with a closed-form recalibration.
258. Diagnostics for Individual-Level Prediction Instability in Machine Learning for Healthcare
Core Problem: Overparameterized machine learning models in healthcare exhibit substantial, often undetected, individual-level prediction variability due to optimization and initialization randomness, which can undermine clinical trust and alter treatment decisions, a problem obscured by aggregate performance metrics.
Key Innovation: Proposed an evaluation framework that quantifies individual-level prediction instability using two diagnostics: empirical prediction interval width (ePIW) for continuous risk estimates and empirical decision flip rate (eDFR) for threshold-based decisions, demonstrating that this instability can be significant and impact clinical recommendations.
259. A Case Study on Concept Induction for Neuron-Level Interpretability in CNN
Core Problem: The internal semantics of hidden neurons in Deep Neural Networks (DNNs) remain poorly understood, hindering interpretability in applications like healthcare and autonomous systems.
Key Innovation: Demonstrated that a Concept Induction-based framework for hidden neuron analysis, previously shown effective on ADE20K, successfully generalizes to the SUN2012 dataset, confirming its broader applicability for assigning interpretable semantic labels to neurons.
260. AdURA-Net: Adaptive Uncertainty and Region-Aware Network
Core Problem: Clinical decision-making is challenged by uncertainty in radiology reports and limitations of automated label extraction in complex multilabel datasets, where models are often forced to provide confident predictions without sufficient evidence, posing high risks.
Key Innovation: Proposed AdURA-Net, a geometry-driven adaptive uncertainty-aware framework for reliable thoracic disease classification, featuring adaptive dilated convolution and multiscale deformable alignment, coupled with a Dual Head Loss combining masked binary cross entropy with logit and a Dirichlet evidential learning objective.
261. Ozone Cues Mitigate Reflected Downwelling Radiance in LWIR Absorption-Based Ranging
Core Problem: Passive long-wave infrared (LWIR) absorption-based ranging suffers from significant inaccuracies due to reflected downwelling radiance, which has often been assumed negligible.
Key Innovation: Introduces two new ranging methods (quadspectral and hyperspectral) that utilize characteristic features from ozone absorption to estimate and mitigate the contribution of reflected downwelling radiance, significantly improving ranging accuracy.
262. Diffusion-Based Low-Light Image Enhancement with Color and Luminance Priors
Core Problem: Low-light images suffer from low contrast, noise, and color distortion, which degrades visual quality and impairs the performance of downstream vision tasks, including those potentially used in remote sensing for geohazards.
Key Innovation: Proposes a novel conditional diffusion framework for low-light image enhancement that incorporates a Structured Control Embedding Module (SCEM) to decompose images into informative components (illumination, features, shadow, color cues), guiding the diffusion model for state-of-the-art enhancement and strong generalization.
263. Detecting Transportation Mode Using Dense Smartphone GPS Trajectories and Transformer Models
Core Problem: Accurately detecting transportation modes from dense smartphone GPS trajectories is a key challenge in GeoAI and transportation research, requiring robust models that can handle complex environments and exhibit flexibility in transfer learning.
Key Innovation: Introduces SpeedTransformer, a novel Transformer-based model that relies solely on speed inputs from dense smartphone GPS trajectories to infer transportation modes, outperforming traditional deep learning models and demonstrating strong flexibility in transfer learning across geographical regions.
264. TENG-BC: Unified Time-Evolving Natural Gradient for Neural PDE Solvers with General Boundary Conditions
Core Problem: Accurately solving time-dependent partial differential equations (PDEs) with neural networks is challenging due to long-time error accumulation and the difficulty of effectively enforcing general boundary conditions (Dirichlet, Neumann, Robin, mixed types).
Key Innovation: TENG-BC, a high-precision neural PDE solver based on the Time-Evolving Natural Gradient. It performs a boundary-aware optimization at each time step that jointly enforces interior dynamics and various boundary conditions within a unified framework, admitting a natural-gradient interpretation for stable time evolution without delicate penalty tuning.
265. Improved Adversarial Diffusion Compression for Real-World Video Super-Resolution
Core Problem: Existing diffusion models for real-world video super-resolution are slow due to multi-step sampling, and one-step networks are still heavy. Direct adversarial diffusion compression (ADC) fails to balance spatial details and temporal consistency.
Key Innovation: An improved ADC method for Real-VSR that distills a large diffusion Transformer into a pruned 2D Stable Diffusion-based backbone augmented with lightweight 1D temporal convolutions, and introduces a dual-head adversarial distillation scheme for disentangled discrimination of details and consistency.
266. DreamWorld: Unified World Modeling in Video Generation
Core Problem: Existing video generation models lack a coherent and unified understanding of the world, often incorporating only single forms of knowledge or relying on rigid alignment, leading to visual instability and temporal flickering when modeling multiple heterogeneous dimensions.
Key Innovation: DreamWorld, a unified framework that integrates complementary world knowledge into video generators via a Joint World Modeling Paradigm, jointly predicting video pixels and features, and employing Consistent Constraint Annealing (CCA) and Multi-Source Inner-Guidance to mitigate instability and enforce learned priors.
267. Random Wins All: Rethinking Grouping Strategies for Vision Tokens
Core Problem: Vision Transformers suffer from quadratic complexity, and various complex grouping strategies have been proposed without a clear understanding of whether simpler, more unified methods could be equally or more effective.
Key Innovation: Proposes a simple random grouping strategy for vision tokens that empirically outperforms many carefully designed methods across visual tasks, point cloud processing, and vision-language models, identifying crucial elements for effective grouping strategies.
268. ArtiFixer: Enhancing and Extending 3D Reconstruction with Auto-Regressive Diffusion Models
Core Problem: Per-scene 3D reconstruction methods extrapolate poorly to under-observed areas, and existing generative prior methods for artifact correction lack scalability and produce inconsistent or low-quality outputs in unseen regions.
Key Innovation: Proposes a two-stage pipeline leveraging a powerful bidirectional generative model with a novel opacity mixing strategy for consistency and extrapolation, distilled into a causal auto-regressive model that generates hundreds of frames in a single pass, significantly improving 3D reconstruction quality and extending to previously unobserved regions.
269. COG: Confidence-aware Optimal Geometric Correspondence for Unsupervised Single-reference Novel Object Pose Estimation
Core Problem: Estimating the 6DoF pose of a novel object with a single reference view is challenging due to occlusions, viewpoint changes, and outliers, with existing methods relying on non-differentiable, sparse one-to-one matching.
Key Innovation: Proposes COG, an unsupervised framework that formulates correspondence estimation as a confidence-aware optimal transport problem, producing balanced soft correspondences by predicting point-wise confidences and integrating semantic priors from vision foundation models, achieving comparable or superior performance to supervised methods.
270. Adaptive Dynamic Dehazing via Instruction-Driven and Task-Feedback Closed-Loop Optimization for Diverse Downstream Task Adaptation
Core Problem: Haze removal needs to be adaptive to specific downstream task requirements without retraining, and current methods lack this flexibility.
Key Innovation: A novel adaptive dynamic dehazing framework with a closed-loop optimization mechanism, integrating a task feedback loop and a text instruction interface for real-time, task-adaptive dehazing.
271. Decoupling Stability and Plasticity for Multi-Modal Test-Time Adaptation
Core Problem: Adapting pretrained multi-modal models to evolving test-time distributions often leads to negative transfer in unbiased modalities and catastrophic forgetting in biased modalities.
Key Innovation: DASP (Decoupling Adaptation for Stability and Plasticity), a diagnose-then-mitigate framework that identifies the biased modality based on interdimensional redundancy and employs an asymmetric adaptation strategy with a decoupled architecture, where modality-specific adapters are divided into stable and plastic components, updated differently for biased vs. unbiased modalities.
272. AlignVAR: Towards Globally Consistent Visual Autoregression for Image Super-Resolution
Core Problem: Visual autoregressive (VAR) models for image super-resolution (ISR) face challenges from locality-biased attention, which fragments spatial structures, and residual-only supervision, which accumulates errors across scales, severely compromising global consistency of reconstructed images.
Key Innovation: Proposes AlignVAR, a globally consistent visual autoregressive framework for ISR, featuring Spatial Consistency Autoregression (SCA) to mitigate locality bias and Hierarchical Consistency Constraint (HCC) to stabilize coarse-to-fine refinement, enhancing structural coherence and perceptual fidelity while being efficient.
273. Learning to Explore: Policy-Guided Outlier Synthesis for Graph Out-of-Distribution Detection
Core Problem: Unsupervised graph-level Out-of-Distribution (OOD) detection suffers from incomplete feature space characterization and weak decision boundaries because models are typically trained using only in-distribution data, and existing outlier synthesis approaches rely on fixed, non-adaptive sampling heuristics.
Key Innovation: Proposes a Policy-Guided Outlier Synthesis (PGOS) framework that replaces static heuristics with a learned exploration strategy, training a reinforcement learning agent to navigate low-density regions in a structured latent space and synthesize informative OOD graphs to improve detector robustness.
274. Linking Modality Isolation in Heterogeneous Collaborative Perception
Core Problem: Heterogeneity across agents introduces domain gaps that hinder collaborative perception, further exacerbated by 'modality isolation' where multiple agents with different modalities never co-occur in any training data frame, making existing alignment methods ineffective.
Key Innovation: Proposes CodeAlign, the first efficient, co-occurrence-free alignment framework that smoothly aligns modalities via cross-modal feature-code-feature (FCF) translation, explicitly identifying representation consistency through codebooks and learning direct mappings between modality-specific feature spaces.
275. CoLC: Communication-Efficient Collaborative Perception with LiDAR Completion
Core Problem: Early fusion in collaborative perception for autonomous agents offers high perceptual complementarity but is limited by high communication costs, pushing towards less effective intermediate/late fusion.
Key Innovation: Proposes CoLC, a communication-efficient early collaborative perception framework incorporating LiDAR Completion, Foreground-Aware Point Sampling, Completion-Enhanced Early Fusion, and Dense-Guided Dual Alignment to restore scene completeness under sparse transmission.
276. Bi-cLSTM: Residual-Corrected Bidirectional LSTM for Aero-Engine RUL Estimation
Core Problem: Existing deep learning approaches for Remaining Useful Life (RUL) prediction in safety-critical systems like aero-engines, particularly LSTM-based models, struggle to generalize across varying operating conditions and are sensitive to noise in multivariate sensor data.
Key Innovation: Proposes Bi-cLSTM, a novel Bidirectional Residual Corrected LSTM model that combines bidirectional temporal modeling with an adaptive residual correction mechanism and a condition-aware preprocessing pipeline for robust RUL estimation, outperforming baselines on the NASA C-MAPSS dataset.
277. MME: Mixture of Mesh Experts with Random Walk Transformer Gating
Core Problem: Existing mesh analysis methods often excel on different object classes, lacking a unified framework to leverage their complementary strengths for tasks like classification, retrieval, and semantic segmentation.
Key Innovation: Presents MME, a novel Mixture of Experts (MoE) framework for mesh analysis with a new gate architecture that uses random walks and attention to encourage expert specialization, combined with a dynamic loss balancing scheme, achieving state-of-the-art results in mesh classification, retrieval, and semantic segmentation.
278. Neural Discrimination-Prompted Transformers for Efficient UHD Image Restoration and Enhancement
Core Problem: Efficiently restoring and enhancing Ultra-High-Definition (UHD) images, particularly in challenging conditions like low-light, haze, and blur, by effectively leveraging the implicit neural differences between high-resolution and low-resolution features.
Key Innovation: Proposes UHDPromer, a neural discrimination-prompted Transformer that integrates Neural Discrimination Priors (NDP) into its attention and network mechanisms to globally perceive and selectively use discrimination information, combined with a super-resolution-guided reconstruction approach, achieving state-of-the-art performance with high computational efficiency for UHD image restoration and enhancement.
279. Probabilistic Learning and Generation in Deep Sequence Models
Core Problem: Deep sequence models (DSMs) lack uncertainty awareness, which is crucial for reliable deployment, and exact Bayesian inference for these models is computationally infeasible, while approximate methods face challenges in prior specification and approximation quality.
Key Innovation: Leverages inductive biases in DSMs to design probabilistic inference or structures, developing approximate Bayesian inference for Transformers and HiPPOs, and exploring self-supervision for sequential latent states in generative models to bridge the gap between DSMs and probabilistic models.
280. Principled Fast and Meta Knowledge Learners for Continual Reinforcement Learning
Core Problem: Continual Reinforcement Learning (RL) problems struggle with efficient knowledge transfer and integration across tasks while minimizing catastrophic forgetting of previously learned knowledge.
Key Innovation: Proposes a dual-learner framework for continual RL, comprising a fast learner for knowledge transfer and a meta learner for knowledge integration that explicitly minimizes catastrophic forgetting, along with an adaptive meta warm-up mechanism for rapid adaptation.
281. ShiftLUT: Spatial Shift Enhanced Look-Up Tables for Efficient Image Restoration
Core Problem: Existing Look-Up Table (LUT) based image restoration methods improve performance by expanding the receptive field but introduce significant computational and storage overhead, hindering deployment on edge devices.
Key Innovation: Proposes ShiftLUT, an efficient framework for image restoration that achieves a large receptive field with high efficiency through a Learnable Spatial Shift module, an asymmetric dual-branch architecture, and an Error-bounded Adaptive Sampling (EAS) feature-level LUT compression strategy.
282. Decoupling Motion and Geometry in 4D Gaussian Splatting
Core Problem: Existing 4D Gaussian Splatting methods couple Gaussian motion and geometric attributes, limiting expressiveness for complex motions and often leading to visual artifacts in high-fidelity dynamic scene reconstruction.
Key Innovation: Proposing VeGaS, a novel velocity-based 4D Gaussian Splatting framework that decouples motion and geometry by introducing a Galilean shearing matrix for time-varying velocity and a Geometric Deformation Network to refine Gaussian shapes, achieving state-of-the-art performance in dynamic scene reconstruction.
283. Intent-Context Synergy Reinforcement Learning for Autonomous UAV Decision-Making in Air Combat
Core Problem: Autonomous UAV infiltration in dynamic contested environments remains challenging due to partial observability, conflicting objectives (efficiency vs. survivability), and myopic decision-making in traditional Reinforcement Learning (RL) approaches.
Key Innovation: Proposing the Intent-Context Synergy Reinforcement Learning (ICS-RL) framework, which includes an LSTM-based Intent Prediction Module for proactive planning and a Context-Analysis Synergy Mechanism that decomposes missions into hierarchical sub-tasks with specialized Dueling DQN agents, integrated by a dynamic switching controller, leading to significantly improved mission success and reduced exposure in air combat simulations.
284. Event-Anchored Frame Selection for Effective Long-Video Understanding
Core Problem: Massive frame redundancy and limited context windows make efficient frame selection crucial for long-video understanding with large vision-language models (LVLMs), but prevailing approaches use flat sampling, treating videos as unstructured collections of frames.
Key Innovation: Introducing Event-Anchored Frame Selection (EFS), a hierarchical, event-aware pipeline that first partitions video into visually homogeneous temporal segments using DINO embeddings, then selects query-relevant frames as anchors within each event, and finally refines the keyframe set using an adaptive Maximal Marginal Relevance (MMR) scheme, leading to substantial gains in accuracy for off-the-shelf LVLMs on challenging video understanding benchmarks.
285. A level-wise training scheme for learning neural multigrid smoothers with application to integral equations
Core Problem: Classical multigrid methods fail to effectively solve large and ill-conditioned linear systems derived from convolution-type integral equations because their conventional relaxation smoothers are ineffective at reducing high-frequency error components.
Key Innovation: A novel neural multigrid scheme that replaces classical smoothers with learned neural operators. These operators are trained offline using level-wise loss functions incorporating spectral filtering, allowing them to generalize to new problems and efficiently solve integral equations and potentially PDE problems.
286. Beyond Global Similarity: Towards Fine-Grained, Multi-Condition Multimodal Retrieval
Core Problem: Existing multimodal retrieval benchmarks primarily focus on coarse-grained or single-condition alignment, failing to address real-world scenarios that require fine-grained, multi-condition retrieval based on multiple interdependent constraints across modalities.
Key Innovation: MCMR (Multi-Conditional Multimodal Retrieval), a large-scale benchmark designed to evaluate fine-grained, multi-condition cross-modal retrieval under natural-language queries across five product domains. It assesses the condition-aware reasoning abilities of MLLM-based multimodal retrievers and vision-language rerankers.
287. HeroGS: Hierarchical Guidance for Robust 3D Gaussian Splatting under Sparse Views
Core Problem: 3D Gaussian Splatting (3DGS) performs poorly in novel view synthesis under sparse camera coverage, leading to irregular Gaussian distributions, sparse coverage, blurred backgrounds, and distorted high-frequency areas.
Key Innovation: HeroGS, a hierarchical guidance framework (image, feature, and parameter levels), converts sparse supervision into pseudo-dense guidance, refines high-frequency details, and ensures geometric consistency, significantly improving 3D reconstruction quality under sparse-view conditions.
288. DeAR: Fine-Grained VLM Adaptation by Decomposing Attention Head Roles
Core Problem: Prompt learning for Vision-Language Model (VLM) adaptation often suffers from uncontrolled interactions between learnable and original tokens, leading to a trade-off between task adaptation and preserving zero-shot generalization due to a simplistic layer-centric view.
Key Innovation: DeAR, a framework that achieves fine-grained VLM adaptation by decomposing attention head roles (Attribute, Generalization, Mixed) using a novel Concept Entropy metric. It introduces specialized attribute tokens and a Role-Based Attention Mask to control information flow, improving both task adaptation and generalization.
289. Predictive Reasoning with Augmented Anomaly Contrastive Learning for Compositional Visual Relations
Core Problem: Compositional visual relations (CVR) tasks, which involve identifying an outlier image based on complex compositional rules, remain relatively unexplored and challenging to model.
Key Innovation: PR-A$^2$CL (Predictive Reasoning with Augmented Anomaly Contrastive Learning) uses Augmented Anomaly Contrastive Learning to distil discriminative features and a predict-and-verify paradigm with Predictive Anomaly Reasoning Blocks (PARBs) to iteratively reason about and pinpoint discrepancies in compositional visual relations, significantly outperforming state-of-the-art models.
290. Teacher-Guided Causal Interventions for Image Denoising: Orthogonal Content-Noise Disentanglement in Vision Transformers
Core Problem: Conventional image denoising models struggle to reliably distinguish subtle textures from stochastic noise and learn spurious correlations, leading to over-removed details or residual noise artifacts and reduced robustness under distribution shifts.
Key Innovation: TCD-Net (Teacher-Guided Causal Disentanglement Network) uses causal intervention within a Vision Transformer framework to explicitly decompose content and noise. It integrates Environmental Bias Adjustment, a dual-branch disentanglement head with orthogonality constraints, and a causal prior guided by an AI image generation model, outperforming mainstream methods in fidelity and efficiency.
291. RnG: A Unified Transformer for Complete 3D Modeling from Partial Observations
Core Problem: Existing feed-forward generalizable 3D reconstruction models from sparse images often confine their representations to observed regions, leaving unseen geometry un-modeled.
Key Innovation: Presents RnG (Reconstruction and Generation), a novel feed-forward Transformer that unifies 3D reconstruction and generation by predicting an implicit, complete 3D representation. It proposes a reconstruction-guided causal attention mechanism and treats the KV-cache as an implicit 3D representation, achieving state-of-the-art performance in both tasks.
292. CoSMo3D: Open-World Promptable 3D Semantic Part Segmentation through LLM-Guided Canonical Spatial Modeling
Core Problem: Open-world promptable 3D semantic segmentation remains brittle because semantics are inferred in input sensor coordinates, unlike human perception which interprets parts via functional roles in a canonical space.
Key Innovation: Proposes CoSMo3D, which achieves canonical space perception by inducing a latent canonical reference frame learned directly from data. It creates a unified canonical dataset through LLM-guided alignment and realizes canonicality via a dual-branch architecture, leading to more stable and transferable part semantics and establishing new state of the art.
293. Cross-Modal Guidance for Fast Diffusion-Based Computed Tomography
Core Problem: Obtaining high-quality reconstructions from sparse data sets in applications like neutron CT is challenging, and leveraging complementary imaging modalities typically requires retraining diffusion models with large datasets.
Key Innovation: Proposing a method to incorporate an additional imaging modality (e.g., X-ray CT for neutron CT) without retraining the diffusion prior, enabling accelerated imaging of costly modalities and demonstrating substantial reconstruction quality improvement even with imperfect side modalities.
294. TIMI: Training-Free Image-to-3D Multi-Instance Generation with Spatial Fidelity
Core Problem: Existing Image-to-3D multi-instance generation methods struggle to guarantee precise spatial fidelity and incur substantial training overhead, despite pre-trained models possessing meaningful spatial priors.
Key Innovation: Proposes TIMI, a training-free framework that achieves high spatial fidelity through an Instance-aware Separation Guidance (ISG) module for disentanglement and a Spatial-stabilized Geometry-adaptive Update (SGU) module for geometric preservation, without additional training and with faster inference.
295. Token Reduction via Local and Global Contexts Optimization for Efficient Video Large Language Models
Core Problem: Video Large Language Models (VLLMs) suffer from inefficiency due to redundant visual tokens, and existing pruning methods are suboptimal in spatiotemporal reduction and context preservation.
Key Innovation: Proposes AOT (token Anchors via local-global Optimal Transport), a training-free framework that establishes local and global token anchors within and across frames to comprehensively aggregate informative contexts, achieving efficient token reduction and competitive performance in VLLMs.
296. Autoregressive Synthesis of Sparse and Semi-Structured Mixed-Type Data
Core Problem: Existing synthetic data generation methods assume dense, fixed-schema tabular data, failing to handle sparse, semi-structured formats like JSON, which require inefficient flattening and imputation.
Key Innovation: Origami, an autoregressive transformer-based architecture, tokenizes data records (including nested objects and variable length arrays) into sequences of key, value, and structural tokens, natively handling sparsity, mixed types, and hierarchical structure without flattening or imputation, outperforming baselines on fidelity and utility.
297. FACE: A Face-based Autoregressive Representation for High-Fidelity and Efficient Mesh Generation
Core Problem: Autoregressive models for 3D mesh generation are computationally expensive and inefficient due to flattening meshes into long vertex-coordinate sequences, hindering high-fidelity geometry synthesis.
Key Innovation: FACE, an Autoregressive Autoencoder (ARAE) framework that generates meshes at the face level (one-face-one-token strategy), significantly reducing sequence length and computational cost while maintaining state-of-the-art reconstruction quality.
298. Better Matching, Less Forgetting: A Quality-Guided Matcher for Transformer-based Incremental Object Detection
Core Problem: Catastrophic forgetting in Transformer-based Incremental Object Detection (IOD) due to "background foregrounding," where the Hungarian matcher forces erroneous assignments of background regions to foreground classes, disrupting learned representations.
Key Innovation: A Quality-guided Min-Cost Max-Flow (Q-MCMF) matcher that prunes implausible matches based on geometric quality, avoiding forced assignments and eliminating harmful supervision from background foregrounding, thereby improving IOD performance.
299. Scalable Multi-Task Low-Rank Model Adaptation
Core Problem: Catastrophic performance degradation when scaling multi-task low-rank adaptation (LoRA) to a large number of tasks, caused by parameter and representation misalignment and a fundamental trade-off between regularization and feature discrimination.
Key Innovation: mtLoRA, a scalable solution with Spectral-Aware Regularization, Block-Level Adaptation, and Fine-Grained Routing, which addresses inter-task conflict and improves parameter efficiency, leading to significant performance gains on large-scale multi-task benchmarks.
300. TopoMaskV3: 3D Mask Head with Dense Offset and Height Predictions for Road Topology Understanding
Core Problem: Prior mask-based road topology methods were limited to 2D, suffered from discretization artifacts, and lacked robust standalone 3D prediction. Additionally, geographic data leakage in evaluation was a significant issue.
Key Innovation: Introduces TopoMaskV3, which advances road topology understanding into a robust, standalone 3D predictor via novel dense offset and height prediction heads. It also addresses geographic data leakage by introducing geographically distinct splits and a long-range benchmark for evaluation.
301. InterCoG: Towards Spatially Precise Image Editing with Interleaved Chain-of-Grounding Reasoning
Core Problem: Emerging unified editing models struggle with fine-grained editing in complex multi-entity scenes, particularly when targets are not visually salient and require spatial reasoning.
Key Innovation: Proposes InterCoG, a novel text-vision Interleaved Chain-of-Grounding reasoning framework that first performs object position reasoning in text, then conducts visual grounding with bounding boxes/masks, and finally rewrites the editing description. It includes auxiliary training modules and a new dataset (GroundEdit-45K) for spatial localization accuracy and reasoning interpretability.
302. Transform-Invariant Generative Ray Path Sampling for Efficient Radio Propagation Modeling
Core Problem: Ray tracing for radio propagation modeling suffers from exponential computational complexity, limiting its use in large-scale or real-time applications, and applying generative models faces challenges like sparse rewards and convergence failures in complex environments.
Key Innovation: Proposes a machine-learning-assisted framework using Generative Flow Networks for intelligent ray path sampling, incorporating an experience replay buffer, a uniform exploratory policy, and a physics-based action masking strategy to achieve substantial speedups (up to 1000x on CPU) over exhaustive search while maintaining high coverage accuracy.
303. CoopDiff: A Diffusion-Guided Approach for Cooperation under Corruptions
Core Problem: Cooperative perception in multi-agent systems lacks robustness and generalization due to diverse and unpredictable corruptions in real-world scenarios.
Key Innovation: CoopDiff, a diffusion-based cooperative perception framework that mitigates corruptions via a denoising mechanism, employing a teacher-student paradigm with quality-aware fusion and an ego-guided cross-attention mechanism for robust scene understanding under degradation.
304. Towards Principled Dataset Distillation: A Spectral Distribution Perspective
Core Problem: Existing dataset distillation (DD) methods exhibit substantial performance degradation on long-tailed datasets due to heuristic distribution discrepancy measures and uniform treatment of imbalanced classes.
Key Innovation: Class-Aware Spectral Distribution Matching (CSDM), which reformulates distribution alignment via the spectrum of a kernel function (Spectral Distribution Distance - SDD) and performs amplitude-phase decomposition to adaptively prioritize realism in tail classes, significantly improving performance and stability on long-tailed datasets.
305. Learning Domain-Aware Task Prompt Representations for Multi-Domain All-in-One Image Restoration
Core Problem: Existing all-in-one image restoration (AiOIR) methods typically focus on a specific image domain, limiting their applicability across diverse domains like natural scenes, medical imaging, or remote sensing.
Key Innovation: DATPRL-IR, the first multi-domain all-in-one image restoration method, uses Domain-Aware Task Prompt Representation Learning. It constructs task and domain prompt pools, uses a prompt composition mechanism to create instance-level task and domain representations, and fuses them to guide the restoration process across multiple domains, including remote sensing.
306. Practical Deep Heteroskedastic Regression
Core Problem: Training deep heteroskedastic regression models for uncertainty quantification faces practical challenges like optimization difficulties, representation collapse, and variance overfitting, hindering the trade-off between UQ and mean prediction.
Key Innovation: Proposes a simple and efficient post-hoc procedure to fit a variance model across intermediate layers of a pretrained network on a hold-out dataset, addressing training challenges and achieving on-par or state-of-the-art UQ without compromising mean prediction.
307. Efficient Test-Time Optimization for Depth Completion via Low-Rank Decoder Adaptation
Core Problem: Existing zero-shot depth completion methods, whether diffusion-based or visual-prompt-based, are computationally expensive and slow due to iterative denoising or repeated full network passes during test-time optimization.
Key Innovation: Proposes a lightweight test-time adaptation method for depth completion that updates only a low-dimensional decoder subspace using sparse depth supervision, based on the insight that depth foundation models concentrate depth-relevant information there, achieving state-of-the-art accuracy and efficiency.
308. Downstream Task Inspired Underwater Image Enhancement: A Perception-Aware Study from Dataset Construction to Network Design
Core Problem: Existing underwater image enhancement (UIE) methods primarily focus on human visual perception, failing to reconstruct high-frequency details crucial for downstream computer vision tasks like semantic segmentation and object detection in challenging underwater environments.
Key Innovation: Proposes DTI-UIE framework, an efficient two-branch network with a task-aware attention module, multi-stage training, and a task-driven perceptual loss. It also constructs a Task-Inspired UIE Dataset (TI-UIED) to significantly improve task performance for downstream underwater vision tasks.
309. Phase-Type Variational Autoencoders for Heavy-Tailed Data
Core Problem: Standard Variational Autoencoders (VAEs) struggle to capture heavy-tailed behavior in real-world data due to simple decoder distributions, while existing heavy-tail extensions are limited to predefined parametric families with fixed tail behavior.
Key Innovation: Proposes the Phase-Type Variational Autoencoder (PH-VAE), which uses a latent-conditioned Phase-Type (PH) distribution as its decoder. This flexible and analytically tractable decoder, defined as the absorption time of a continuous-time Markov chain, adapts its tail behavior directly from data, accurately recovering diverse heavy-tailed distributions and capturing cross-dimensional tail dependence.
310. Neural Operator-Grounded Continuous Tensor Function Representation and Its Applications
Core Problem: Current continuous tensor function representations are limited by discrete and linear mode-n products, preventing a truly continuous and nonlinear representation of real-world data and leading to discretization artifacts.
Key Innovation: Proposes Neural Operator-Grounded Continuous Tensor Function Representation (NO-CTR), which uses neural operator-grounded mode-n operators as continuous and nonlinear alternatives to discrete mode-n products. This allows direct mapping from continuous core tensor functions to continuous target tensor functions, providing a more faithful representation of complex real-world data and demonstrating superiority in multi-dimensional data completion tasks.
311. Diagnosing Generalization Failures from Representational Geometry Markers
Core Problem: Anticipating unseen generalization failures in AI models remains a central challenge, with conventional mechanistic approaches struggling to provide high-level predictive signals.
Key Innovation: Proposes a "top-down" approach using representational geometry markers (effective manifold dimensionality and utility) to consistently forecast poor out-of-distribution generalization across diverse AI architectures and datasets, offering robust guidance for model selection.
312. Resolving Blind Inverse Problems under Dynamic Range Compression via Structured Forward Operator Modeling
Core Problem: Recovering radiometric fidelity from unknown dynamic range compression (UDRC) is a challenging blind inverse problem due to unknown forward models and irreversible information loss.
Key Innovation: Introduces the cascaded monotonic Bernstein (CaMB) operator to parameterize the unknown forward model by enforcing monotonicity, and integrates it with a plug-and-play diffusion framework (CaMB-Diff) to robustly recover radiometric fidelity in zero-shot UDRC tasks.
313. Explanation-Guided Adversarial Training for Robust and Interpretable Models
Core Problem: Deep neural networks often lack interpretability and robustness, making them vulnerable to imperceptible perturbations (adversarial attacks) and out-of-distribution data, even when achieving high performance on benign inputs.
Key Innovation: Proposes Explanation-Guided Adversarial Training (EGAT), a unified framework that integrates adversarial training (AT) and explanation-guided learning (EGL). EGAT generates adversarial examples while imposing explanation-based constraints, jointly optimizing classification performance, adversarial robustness, and attributional stability, leading to more robust and interpretable models.
314. BAED: a New Paradigm for Few-shot Graph Learning with Explanation in the Loop
Core Problem: Few-shot graph learning (FSGL) suffers from insufficient labeled data, leading to compromises in model robustness and interpretability, and potential overfitting to noise.
Key Innovation: Introduces BAED, the first explanation-in-the-loop framework for FSGL. It uses belief propagation for label augmentation and an auxiliary graph neural network with gradient backpropagation to extract explanatory subgraphs, improving prediction accuracy, training efficiency, and explanation quality by focusing on informative subgraphs.
315. CoVAE: correlated multimodal generative modeling
Core Problem: Existing Multimodal Variational Autoencoders (MVAEs) rely on fusion strategies in latent space that destroy the joint statistical structure of multimodal data, impacting generation and uncertainty quantification.
Key Innovation: Introduces Correlated Variational Autoencoders (CoVAE), a new generative architecture that explicitly captures the correlations between modalities, demonstrating accurate cross-modal reconstruction and effective quantification of associated uncertainties.
316. Event-Only Drone Trajectory Forecasting with RPM-Modulated Kalman Filtering
Core Problem: The use of high-temporal-resolution event cameras for drone trajectory prediction remains limited, especially for robust and accurate short- to medium-horizon forecasting without reliance on RGB imagery or extensive training data.
Key Innovation: Introduces an event-only drone forecasting method that exploits propeller-induced motion cues extracted directly from raw event data and fuses them within an RPM-aware Kalman filtering framework, demonstrating robust and accurate trajectory forecasting.
317. Revealing Combinatorial Reasoning of GNNs via Graph Concept Bottleneck Layer
Core Problem: The combinatorial reasoning underlying Graph Neural Network (GNN) predictions is often hidden within their black-box architectures, and existing interpretability methods are either post-hoc approximations or only uncover hard logical rules.
Key Innovation: A graph concept bottleneck layer that integrates into any GNN architecture, guiding it to predict discriminative global graph concepts and projecting these scores to class labels via a sparse linear layer, thereby enforcing soft logical rules and quantifying concept contributions for improved interpretability.
318. Latent attention on masked patches for flow reconstruction
Core Problem: Vision transformers have limited adoption in scientific disciplines like fluid dynamics, and there's a need for interpretable and accurate masked flow reconstruction methods.
Key Innovation: Latent Attention on Masked Patches (LAMP), an interpretable regression-based modified vision transformer that accurately reconstructs full flow fields from highly masked and noisy inputs in fluid dynamics, and yields physically interpretable optimal sensor-placement maps.
319. FluxMem: Adaptive Hierarchical Memory for Streaming Video Understanding
Core Problem: Efficiently processing and understanding streaming video data in real-time, particularly by compressing redundant visual memory, to overcome latency and memory constraints while maintaining accuracy.
Key Innovation: Introduces FluxMem, a training-free framework for streaming video understanding that uses a hierarchical, two-stage memory compression (Temporal Adjacency Selection and Spatial Domain Consolidation) with a self-adaptive token compression mechanism, achieving state-of-the-art results with reduced latency and memory.
320. A 3D mesh convolution-based autoencoder for geometry compression
Core Problem: Efficiently compressing irregular 3D mesh data while preserving geometry and connectivity, without requiring specific preprocessing or manifold/watertightness conditions, which are limitations of existing methods.
Key Innovation: Introduces a novel 3D mesh convolution-based autoencoder that learns features directly from mesh faces, preserves connectivity through dedicated pooling/unpooling operations, and outperforms state-of-the-art methods in 3D mesh geometry reconstruction and latent space classification.
321. Adaptive Confidence Regularization for Multimodal Failure Detection
Core Problem: The largely unexplored problem of failure detection in multimodal contexts, where models need to be reliable in high-stakes domains but often suffer from confidence degradation during failures.
Key Innovation: Proposes Adaptive Confidence Regularization (ACR), a novel framework that uses an Adaptive Confidence Loss to penalize confidence degradations and Multimodal Feature Swapping for generating failure-aware training examples, achieving consistent and robust gains in multimodal failure detection.
322. High-Resolution Range Profile Classifiers Require Aspect-Angle Awareness
Core Problem: High-Resolution Range Profile (HRRP) classification often performs suboptimally because prior work assumes incomplete or unavailable aspect-angle information during training or inference.
Key Innovation: Demonstrating that both single-profile and sequential HRRP classifiers consistently benefit from aspect-angle awareness (average 7% accuracy gain), and showing that estimated angles (via a causal Kalman filter) preserve most of these gains, making the approach practical.
323. Pulse-Driven Neural Architecture: Learnable Oscillatory Dynamics for Robust Continuous-Time Sequence Processing
Core Problem: Continuous-time recurrent networks struggle with robustness to input interruptions in sequence processing.
Key Innovation: PDNA (Pulse-Driven Neural Architecture) augments continuous-time recurrent networks with learnable oscillatory dynamics and a self-attend module, significantly improving robustness to input interruptions compared to baselines.
324. KROM: Kernelized Reduced Order Modeling
Core Problem: The need for fast and adaptive solutions for nonlinear partial differential equations, especially for problems with complex, problem-specific structures (e.g., boundary behavior, oscillations, nonsmooth features) where traditional methods or hand-tuned kernels may be insufficient.
Key Innovation: KROM (Kernelized Reduced Order Modeling), a kernel-based framework that formulates PDE solution as a minimum-norm recovery problem, accelerating kernel solves via sparse Cholesky factorization. It uses an empirical kernel constructed from a snapshot library of PDE solutions, allowing it to adapt to problem-specific structures and yield an implicit, localized reduced model, outperforming baselines in nonsmooth regimes.
325. Heterophily-Agnostic Hypergraph Neural Networks with Riemannian Local Exchanger
Core Problem: Traditional Hypergraph Neural Networks (HGNNs) follow the homophily assumption and struggle with prevalent heterophilic hypergraphs, which require effective modeling of long-range dependencies.
Key Innovation: HealHGNN, a novel Heat-Exchanger with Adaptive Locality for Hypergraph Neural Network, achieves heterophily-agnostic message passing through Riemannian geometry. It uses an adaptive local (heat) exchanger to capture rich long-range dependencies via the Robin condition and preserves representation distinguishability via source terms.
326. Random Features for Operator-Valued Kernels: Bridging Kernel Methods and Neural Operators
Core Problem: A lack of rigorous theoretical analysis for the generalization properties of random feature methods, especially when extended to spectral regularization and operator-valued kernels, which are crucial for understanding neural operators and neural networks.
Key Innovation: Extends the analysis of random feature methods to spectral regularization and operator-valued kernels, providing a unified framework for rigorous theoretical analysis of neural operators and neural networks through the Neural Tangent Kernel (NTK), establishing optimal learning rates and minimax rates.
327. Super-resolution of turbulent reacting flows on complex meshes using graph neural networks
Core Problem: State-of-the-art deep learning models for reconstructing small-scale structures in turbulent flows are predominantly restricted to structured uniform meshes, limiting their applicability to data associated with complex geometries (structured non-uniform or unstructured meshes).
Key Innovation: A methodology leveraging the inherent flexibility of Graph Neural Networks (GNNs) with message passing layers to reconstruct unresolved small-scale structures from low-resolution data on complex meshes, demonstrated for reacting channel flow and internal combustion engine simulations.
328. Adaptive-Growth Randomized Neural Networks for Level-Set Computation of Multivalued Nonlinear First-Order PDEs with Hyperbolic Characteristics
Core Problem: Computing multivalued solutions of nonlinear first-order PDEs with hyperbolic characteristics is challenging due to solutions becoming unions of multiple branches after singularity formation and the high dimensionality of level-set formulations.
Key Innovation: Proposes an Adaptive-Growth Randomized Neural Network (AG-RaNN) method combined with an adaptive collocation strategy and a layer-growth mechanism to efficiently recover multivalued structures and resolve nonsmooth features in high-dimensional PDE solutions, with potential applications in seismic wave modeling.
329. Certifiable Estimation with Factor Graphs
Core Problem: Inference in factor graph models for state estimation typically uses local optimization, which can converge to suboptimal solutions, posing reliability concerns, while existing certifiable estimators are computationally costly and require specialized expertise.
Key Innovation: Synthesizes factor graphs and certifiable estimation into a unified framework, demonstrating that factor graph structure is preserved under Shor's relaxation and Burer-Monteiro factorization, enabling certifiable estimation to be implemented using existing factor graph libraries and workflows.
330. Randomized Neural Networks for Partial Differential Equation on Static and Evolving Surfaces
Core Problem: Numerically solving Partial Differential Equations (PDEs) on static and evolving surfaces is challenging due to geometric complexity and the need for mesh updates, while existing neural-network methods can be costly or inaccurate.
Key Innovation: Developing a Randomized Neural Network (RaNN) method for solving PDEs on static and evolving surfaces, where hidden-layer parameters are fixed and output coefficients are efficiently determined, and for evolving surfaces, learning surface evolution via a flow-map representation to solve PDEs on a space-time collocation set, demonstrating broad applicability and favorable accuracy-efficiency.
331. GCTAM: Global and Contextual Truncated Affinity Combined Maximization Model For Unsupervised Graph Anomaly Detection
Core Problem: Existing truncated affinity maximization (TAM) methods for unsupervised graph anomaly detection use rigid thresholds for truncating suspicious nodes, ignoring node specificity and high-order affinities, which limits their effectiveness in identifying anomalies in real-world graphs.
Key Innovation: Proposes GCTAM, a novel truncation model that combines contextual and global affinity maximization to truncate anomalous nodes more effectively. It uses contextual truncation to decrease anomalous node affinity and global truncation to increase normal node affinity, significantly outperforming peer methods on real-world datasets.
332. LOCUS: A Distribution-Free Loss-Quantile Score for Risk-Aware Predictions
Core Problem: Modern machine learning models, despite high average accuracy, can make critical mistakes that dominate deployment costs, and there is a need for per-input reliability scores for risk-aware predictions and control of large-loss events.
Key Innovation: LOCUS, a distribution-free wrapper that produces a per-input loss-scale reliability score for a fixed prediction function, modeling the realized loss and allowing for transparent flagging rules with distribution-free control of large-loss events, yielding effective risk ranking.
333. Instrumental and Proximal Causal Inference with Gaussian Processes
Core Problem: Existing Instrumental Variable (IV) and Proximal Causal Learning (Proxy) methods for causal inference in the presence of unobserved confounding rarely provide reliable epistemic uncertainty (EU) quantification.
Key Innovation: A Deconditional Gaussian Process (DGP) framework for uncertainty-aware causal learning, which recovers popular kernel estimators as the posterior mean, provides principled and well-calibrated EU via posterior variance, and enables systematic model selection through marginal log-likelihood optimization.
334. OmniTracker: Unifying Object Tracking by Tracking-with-Detection
Core Problem: Visual Object Tracking (VOT) tasks are divided into instance tracking and category tracking, leading to divergent solutions, redundant training expenses, and parameter overhead, despite their shared goal of estimating target object positions.
Key Innovation: OmniTracker, a unified tracking model based on a novel tracking-with-detection paradigm, which combines appearance priors for detection with candidate bounding boxes for association, resolving all tracking tasks with a fully shared network architecture, model weights, and inference pipeline, achieving on-par or better results across various datasets.
335. NUBO: A Transparent Python Package for Bayesian Optimization
Core Problem: Providing an accessible, transparent, and user-friendly Bayesian optimization framework for researchers to optimize expensive black-box functions.
Key Innovation: Developing NUBO, an open-source Python package for Bayesian optimization that prioritizes transparency, modularity, and user experience, supporting various optimization scenarios with extensively tested algorithms.
336. Velocity Disambiguation for Video Frame Interpolation
Core Problem: Existing video frame interpolation methods struggle with precise object movement prediction due to "time indexing" leading to blurry frames and motion ambiguity, especially for long-range motion.
Key Innovation: Proposing "distance indexing" to provide explicit hints on object travel distance, and an iterative reference-based estimation strategy to break down long-range predictions, significantly improving perceptual quality in arbitrary time interpolations and enabling novel video editing tasks.
337. Beyond DAGs: A Latent Partial Causal Model for Multimodal Learning
Core Problem: Directed Acyclic Graphs (DAGs) are insufficient for modeling the complexity and heterogeneous generative processes of large-scale multimodal data, which may involve multiple or opposing causal structures.
Key Innovation: Proposing a novel latent partial causal model for multimodal data representation learning with two latent coupled variables, establishing an identifiability result for MMCL, and demonstrating its potential for representation disentanglement and improved domain generalization in pre-trained models like CLIP.
338. Towards Camera Open-set 3D Object Detection for Autonomous Driving Scenarios
Core Problem: Conventional camera-based 3D object detectors in autonomous driving are limited to predefined object sets, posing safety risks when encountering novel objects.
Key Innovation: OS-Det3D, a two-stage framework for camera-based open-set 3D object detection, which uses a 3D object discovery network (ODN3D) with LiDAR cues for class-agnostic proposals and a joint selection (JS) module to filter low-quality proposals, enhancing detection of both known and unknown objects.
339. A Tidal Current Speed Forecasting Model based on Multi-Periodicity Learning
Core Problem: Inaccurate tidal current speed forecasting, primarily due to the complex multi-periodicity of tidal variations, which challenges traditional physical models and hinders high tidal energy penetration.
Key Innovation: The Wavelet-Enhanced Convolutional Network (WECN) framework that learns multi-periodicity by embedding intra-period and inter-period variations into a 2D tensor, integrating time-frequency analysis, and optimizing hyperparameters for enhanced stability and accuracy in tidal current speed forecasting.
340. FiLo++: Zero-/Few-Shot Anomaly Detection by Fused Fine-Grained Descriptions and Deformable Localization
Core Problem: Existing zero-shot and few-shot anomaly detection methods use generic descriptions and struggle to accurately localize anomalies of varying shapes and sizes, especially when extensive normal samples are unavailable.
Key Innovation: FiLo++, a method combining Fused Fine-Grained Descriptions (FusDes) generated by LLMs for task-specific anomaly descriptions, and Deformable Localization (DefLoc) which integrates a vision foundation model with position-enhanced text descriptions and a Multi-scale Deformable Cross-modal Interaction module for accurate localization of diverse anomaly shapes.
341. Benchmarking Self-Supervised Learning Methods for Accelerated MRI Reconstruction
Core Problem: Accelerating MRI reconstruction from undersampled measurements is challenging, and supervised deep learning methods require expensive or impossible-to-obtain fully-sampled ground truth images, while existing self-supervised methods lack systematic comparison.
Key Innovation: Presents SSIBench, a modular benchmarking framework for 18 self-supervised MRI reconstruction methods across diverse scenarios, revealing performance landscapes, and proposing a novel Multi-Operator Equivariant Imaging loss to guide future improvements in GT-free imaging.
342. Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models
Core Problem: Direct Reinforcement Learning (RL) training struggles to activate complex reasoning capabilities in Multimodal Large Language Models (MLLMs) due to the absence of substantial high-quality multimodal reasoning data.
Key Innovation: Vision-R1, a reasoning MLLM that improves multimodal reasoning capability by constructing a high-quality multimodal CoT dataset without human annotations and employing a Progressive Thinking Suppression Training (PTST) strategy with Group Relative Policy Optimization (GRPO), achieving significant improvements on multimodal math reasoning benchmarks.
343. Wasserstein Distances Made Explainable: Insights Into Dataset Shifts and Transport Phenomena
Core Problem: While Wasserstein distances are powerful for comparing data distributions, understanding the specific factors (e.g., data subgroups, input features) that contribute to a high or low distance, indicating dataset shifts or inhomogeneities, remains challenging.
Key Innovation: A novel Explainable AI solution that efficiently and accurately attributes Wasserstein distances to various data components (subgroups, input features, interpretable subspaces), providing clear insights into dataset shifts and transport phenomena, demonstrated across diverse datasets.
344. EquiReg: Equivariance Regularized Diffusion for Inverse Problems
Core Problem: Existing diffusion models for inverse problems rely on isotropic Gaussian approximations, which can push estimates off the data manifold and produce inconsistent, poor reconstructions, especially under reduced sampling.
Key Innovation: Proposes Equivariance Regularized (EquiReg) diffusion, a plug-and-play framework that improves posterior sampling by penalizing trajectories deviating from the data manifold using manifold-preferential equivariant functions. This implicitly accelerates convergence and enables high-quality reconstructions in image restoration and PDE solving.
345. RocketStack: Level-aware Deep Recursive Ensemble Learning Architecture
Core Problem: Deep stacking in ensemble learning is uncommon due to challenges like feature redundancy, complexity, and computational burden, limiting the benefits of integrating multiple base learners at deeper levels.
Key Innovation: Introduces RocketStack, a level-aware recursive stacking architecture that extends to ten levels. It addresses limitations through incremental pruning of weaker learners using out-of-fold scores (regularized by Gaussian perturbations) and periodic feature compression, achieving increased accuracy with depth and sublinear computational growth.
346. Consistency-Driven Calibration and Matching for Few-Shot Class-Incremental Learning
Core Problem: Few-Shot Class Incremental Learning (FSCIL) struggles to balance old and new knowledge due to prototype bias and rigid structures, leading to knowledge conflict and limiting the expressive capacity of the embedding space.
Key Innovation: Proposes the Consistency-driven Calibration and Matching (ConCM) framework to mitigate knowledge conflict in FSCIL. It uses memory-aware prototype calibration to enhance conceptual center consistency and dynamic structure matching to adaptively align features to a session-specific optimal manifold space, achieving state-of-the-art performance.
347. MonoFusion: Sparse-View 4D Reconstruction via Monocular Fusion
Core Problem: Reconstructing dynamic scenes from sparse-view videos is challenging because dense multi-view methods struggle with limited overlap, and existing multi-view setups are prohibitively expensive and cannot capture diverse in-the-wild scenes.
Key Innovation: MonoFusion, a method that carefully aligns independent monocular reconstructions from a small set of sparse-view cameras to produce time- and view-consistent dynamic scene reconstructions, achieving higher quality than prior art, particularly when rendering novel views.
348. Federated Nonlinear System Identification
Core Problem: Identifying parameters of linearly-parameterized nonlinear systems in a federated learning setting, where data is distributed across multiple clients, and ensuring convergence and effectiveness compared to centralized approaches.
Key Innovation: Establishes theoretical guarantees for federated nonlinear system identification, demonstrating improved convergence rates with an increasing number of clients, and experimentally validates the method on nonlinear dynamical systems. This approach could be relevant for identifying parameters of distributed geohazard systems.
349. Time-Aware One Step Diffusion Network for Real-World Image Super-Resolution
Core Problem: Existing diffusion-based real-world image super-resolution methods struggle to fully leverage generative priors in pre-trained stable-diffusion models due to using a fixed timestep, leading to suboptimal performance and a lack of controllable trade-offs.
Key Innovation: Proposes TADSR, a Time-Aware one-step Diffusion Network, which introduces a Time-Aware VAE Encoder and a Time-Aware VSD loss to dynamically vary timesteps and latent features, enabling more effective utilization of SD's generative capabilities and achieving controllable trade-offs between fidelity and realism in super-resolution.
350. UniView: Enhancing Novel View Synthesis From A Single Image By Unifying Reference Features
Core Problem: Synthesizing novel views from a single image is highly ill-posed, leading to severe distortions in unobserved areas due to reliance on ambiguity priors and interpolation.
Key Innovation: Proposes UniView, a novel model that enhances novel view synthesis by unifying reference features from similar objects, leveraging a retrieval and augmentation system with an MLLM for reference selection, a plug-and-play adapter module for dynamic reference feature generation, and a decoupled triple attention mechanism for multi-branch feature integration, significantly improving performance.
351. Beyond Frame-wise Tracking: A Trajectory-based Paradigm for Efficient Point Cloud Tracking
Core Problem: Existing 3D single object tracking methods are either efficient but lack long-term context (two-frame) or robust but computationally expensive (sequence-based), creating a dilemma for real-time applications.
Key Innovation: Introduction of TrajTrack, a lightweight trajectory-based paradigm that enhances two-frame trackers by implicitly learning motion continuity from historical bounding box trajectories, achieving state-of-the-art performance efficiently.
352. Seg2Track-SAM2: SAM2-based Multi-object Tracking and Segmentation
Core Problem: Direct application of foundation models like SAM2 to Multi-Object Tracking and Segmentation (MOTS) is limited by the absence of explicit identity management mechanisms and by growing memory requirements during tracking.
Key Innovation: Introduction of Seg2Track-SAM2, a framework integrating pretrained object detectors with SAM2 and a dedicated Seg2Track module for track initialization, data association, and refinement, achieving high association accuracy and memory efficiency without dataset-specific fine-tuning.
353. GraphUniverse: Synthetic Graph Generation for Evaluating Inductive Generalization
Core Problem: Existing synthetic graph benchmarks are limited to single-graph, transductive settings, hindering systematic evaluation of inductive generalization for graph learning models.
Key Innovation: GraphUniverse, a framework for generating entire families of graphs with persistent semantic communities and controlled structural properties, enables the first systematic evaluation of inductive generalization at scale, revealing that strong transductive performance is a poor predictor of inductive generalization.
354. Quantized Visual Geometry Grounded Transformer
Core Problem: Billion-scale Visual Geometry Grounded Transformers (VGGTs) for 3D reconstruction have prohibitive computational and memory costs, and Post-Training Quantization (PTQ) faces unique obstacles due to heavy-tailed activation distributions and unstable calibration sample selection from multi-view 3D data.
Key Innovation: QuantVGGT, the first quantization framework for VGGTs, introduces Dual-Smoothed Fine-Grained Quantization to mitigate heavy-tailed distributions and Noise-Filtered Diverse Sampling for stable quantization ranges, achieving significant memory reduction and acceleration while maintaining reconstruction accuracy.
355. CubistMerge: Spatial-Preserving Token Merging For Diverse ViT Backbones
Core Problem: Many modern Vision Transformer (ViT) backbones adopt spatial architectural designs, but existing token reduction methods fail to preserve the spatial structure these architectures depend on, limiting efficiency and compatibility.
Key Innovation: CubistMerge, a simple yet effective token merging method, maintains spatial integrity by employing a 2D reduction strategy, a spatial-aware merging algorithm, and a novel max-magnitude-per-dimension token representation, achieving significant speedup with minimal accuracy drop across various vision tasks and ViT architectures.
356. Arbitrary Generative Video Interpolation
Core Problem: Existing generative video frame interpolation (VFI) methods are constrained to synthesize a fixed number of intermediate frames, lacking flexibility for arbitrary frame rates or total sequence duration.
Key Innovation: Introduces ArbInterp, a novel generative VFI framework that enables interpolation at any timestamp and length using Timestamp-aware Rotary Position Embedding (TaRoPE) for fine-grained control and a decoupled appearance-motion conditioning strategy for seamless spatiotemporal transitions.
357. Multi-Marginal Flow Matching with Adversarially Learnt Interpolants
Core Problem: Inferring underlying trajectories and dynamics from sampled observations at discrete time points, especially when ground-truth trajectories are unavailable, is challenging for existing multi-marginal flow matching algorithms.
Key Innovation: Proposes ALI-CFM, a novel multi-marginal flow matching method that uses a GAN-inspired adversarial loss to fit neurally parametrised interpolant curves, yielding smooth and unique trajectories that are then marginalized by a flow matching algorithm to train a vector field for the underlying dynamics.
358. PHyCLIP: $\ell_1$-Product of Hyperbolic Factors Unifies Hierarchy and Compositionality in Vision-Language Representation Learning
Core Problem: Vision-language models struggle to simultaneously express both hierarchical structures within concept families and compositional structures across different concept families.
Key Innovation: PHyCLIP employs an $\ell_1$-Product metric on a Cartesian product of Hyperbolic factors, allowing intra-family hierarchies to emerge within individual hyperbolic factors and cross-family composition to be captured by the $\ell_1$-product metric, leading to improved performance and interpretability.
359. Spotlight on Token Perception for Multimodal Reinforcement Learning
Core Problem: Most existing methods in multimodal Reinforcement Learning with Verifiable Rewards (RLVR) neglect the critical role of visual perception in the optimization process, limiting the reasoning capabilities of Large Vision-Language Models (LVLMs).
Key Innovation: Visually-Perceptive Policy Optimization (VPPO) is a novel policy gradient algorithm that explicitly leverages token perception to refine the learning signal, reweighting trajectory advantages by overall visual dependency and focusing policy updates on perceptually pivotal tokens, significantly enhancing multimodal reasoning.
360. Mono4DGS-HDR: High Dynamic Range 4D Gaussian Splatting from Alternating-exposure Monocular Videos
Core Problem: Reconstructing renderable 4D high dynamic range (HDR) scenes from unposed monocular low dynamic range (LDR) videos is a challenging problem that has not been adequately addressed.
Key Innovation: Introduces Mono4DGS-HDR, the first system for reconstructing renderable 4D HDR scenes from unposed monocular LDR videos using a two-stage Gaussian Splatting optimization. It learns an initial HDR video Gaussian representation in orthographic space, then refines it in world space with camera poses, and includes a temporal luminance regularization for consistency.
361. VoMP: Predicting Volumetric Mechanical Property Fields
Core Problem: Physical simulations rely on spatially-varying mechanical properties that are often laboriously hand-crafted, lacking efficient and accurate prediction methods for 3D objects.
Key Innovation: VoMP, a feed-forward method trained to predict Young's modulus, Poisson's ratio, and density throughout the volume of 3D objects. It aggregates per-voxel multi-view features and uses a Geometry Transformer to predict physically plausible per-voxel material latent codes, outperforming prior art in accuracy and speed.
362. Data-Augmented Deep Learning for Downhole Depth Sensing and Validation
Core Problem: Accurate downhole depth measurement in oil and gas well operations, particularly collar correlation using CCL, is hampered by underdeveloped preprocessing methods and limited availability of real well data for training neural network models.
Key Innovation: A system integrated into a downhole toolstring for CCL log acquisition and comprehensive data augmentation methods (standardization, label distribution smoothing, random cropping, label smoothing regularization, time scaling, multiple sampling). These methods significantly enhance neural network model generalization capabilities for casing collar recognition, achieving improved F1 scores and practical applicability.
363. Transmit Weights, Not Features: Orthogonal-Basis Aided Wireless Point-Cloud Transmission
Core Problem: Efficient and robust wireless transmission of 3D point clouds is challenging, especially in bandwidth-constrained environments, while maintaining geometric fidelity.
Key Innovation: A semantic wireless transmission framework for 3D point clouds built on Deep Joint Source - Channel Coding (DeepJSCC) that transmits combination weights over a receiver-side semantic orthogonal feature pool, enabling compact representations and robust reconstruction with a folding-based decoder.
364. GeoTeacher: Geometry-Guided Semi-Supervised 3D Object Detection
Core Problem: Semi-supervised 3D object detection methods often overlook the model's low sensitivity to object geometries with limited labeled data, making it difficult to capture crucial geometric information for enhancing object perception and localization.
Key Innovation: GeoTeacher, a geometry-guided semi-supervised 3D object detection method that enhances the student model's ability to capture geometric relations. It introduces a keypoint-based geometric relation supervision module and a voxel-wise data augmentation strategy with a distance-decay mechanism, achieving new state-of-the-art results on ONCE and Waymo datasets.
365. Vision-DeepResearch: Incentivizing DeepResearch Capability in Multimodal Large Language Models
Core Problem: Existing multimodal large language models (MLLMs) for multimodal search are naive, limited in reasoning depth and search breadth, and struggle to retrieve key evidence in real-world scenarios with substantial visual noise and complex questions.
Key Innovation: Proposes Vision-DeepResearch, a new multimodal deep-research paradigm that performs multi-turn, multi-entity, and multi-scale visual and textual search to robustly interact with real-world search engines under heavy noise, internalizing deep-research capabilities into the MLLM via cold-start supervision and RL training.
366. Contribution-aware Token Compression for Efficient Video Understanding via Reinforcement Learning
Core Problem: Redundancy in video tokens introduces significant computational overhead for video large language models, and existing compression methods that prioritize high attention scores do not necessarily correlate with actual contribution to correct predictions.
Key Innovation: Proposes CaCoVID, a novel Contribution-aware token Compression algorithm for VIDeo understanding, which uses a reinforcement learning framework to optimize a policy network for selecting token combinations with the greatest contribution to correct predictions, and a combinatorial policy optimization algorithm to accelerate convergence.
367. Vision-DeepResearch Benchmark: Rethinking Visual and Textual Search for Multimodal Large Language Models
Core Problem: Existing benchmarks for evaluating visual and textual search abilities of Vision-DeepResearch systems in MLLMs are limited: they are not visual search-centric (answers leaked via text or MLLM world knowledge) and use overly idealized evaluation scenarios (simple image/text retrieval).
Key Innovation: Constructs the Vision-DeepResearch benchmark (VDR-Bench) with 2,000 VQA instances, created via a multi-stage curation pipeline and expert review, designed to assess systems under realistic real-world conditions. Proposes a simple multi-round cropped-search workflow to improve visual retrieval capabilities in realistic scenarios.
368. Self-Supervised Learning via Flow-Guided Neural Operator on Time-Series Data
Core Problem: Existing self-supervised learning (SSL) methods for time-series data, such as masked autoencoders, rely on reconstructing inputs from a fixed, predetermined masking ratio, limiting flexibility and performance in learning versatile representations.
Key Innovation: Introduces Flow-Guided Neural Operator (FGNO), a novel framework combining operator learning with flow matching for SSL on time-series data, which treats corruption level as a degree of freedom, learns mappings in functional spaces, and extracts a rich hierarchy of features using clean inputs for representation extraction during inference.
369. TS-Haystack: A Multi-Scale Retrieval Benchmark for Time Series Language Models
Core Problem: Long-context retrieval remains a major limitation for Time Series Language Models (TSLMs), as they are typically trained on short sequences, while real-world sensor streams are much longer, requiring precise temporal localization under strict computational constraints.
Key Innovation: TS-Haystack, a long-context temporal retrieval benchmark comprising ten task types across four categories, using controlled needle insertion into longitudinal accelerometer recordings to systematically evaluate TSLMs and highlight the degradation of retrieval performance with increasing context length despite preserved classification accuracy.
370. OmniZip: Learning a Unified and Lightweight Lossless Compressor for Multi-Modal Data
Core Problem: Most learning-based lossless compressors are designed for a single modality, leading to redundant deployments in multi-modal settings, and existing multi-modal large language models are too complex for practical compression.
Key Innovation: Proposing OmniZip, a unified and lightweight lossless compressor for diverse multi-modal data (image, text, speech, etc.), incorporating a modality-unified tokenizer, modality-routing context learning, and modality-routing feedforward design, achieving superior compression efficiency and near real-time inference on edge devices.
371. ToProVAR: Efficient Visual Autoregressive Modeling via Tri-Dimensional Entropy-Aware Semantic Analysis and Sparsity Optimization
Core Problem: Visual Autoregressive (VAR) models, despite enhancing generation quality, face a critical efficiency bottleneck, particularly in later stages of the generation process.
Key Innovation: Proposing ToProVAR, a novel optimization framework for VAR models that leverages attention entropy to characterize semantic projections across different dimensions, uncovering sparsity patterns (token, layer, scale) and applying fine-grained optimization strategies, achieving significant acceleration (up to 3.4x) with minimal quality loss.
372. WISER: Wider Search, Deeper Thinking, and Adaptive Fusion for Training-Free Zero-Shot Composed Image Retrieval
Core Problem: Existing Zero-Shot Composed Image Retrieval (ZS-CIR) methods, relying solely on Text-to-Image (T2I) or Image-to-Image (I2I) paradigms, fail to effectively leverage their complementary strengths under diverse query intents, leading to limitations in fine-grained details or complex semantic modifications.
Key Innovation: Proposing WISER, a training-free framework that unifies T2I and I2I via a 'retrieve-verify-refine' pipeline, performing wider search, adaptive fusion with a verifier, and deeper thinking through structured self-reflection for uncertain retrievals, significantly outperforming previous methods in ZS-CIR.
373. AgentVista: Evaluating Multimodal Agents in Ultra-Challenging Realistic Visual Scenarios
Core Problem: Existing multimodal benchmarks primarily evaluate single-turn visual reasoning or specific tool skills, failing to capture the realism, visual subtlety, and long-horizon tool use required for practical real-world multimodal agents.
Key Innovation: Introducing AgentVista, a benchmark for generalist multimodal agents that spans 25 sub-domains across 7 categories, featuring realistic, detail-rich visual scenarios and natural hybrid tool use, exposing significant gaps in state-of-the-art models' ability for long-horizon multimodal tool use.
374. Theoretical Foundations of Superhypergraph and Plithogenic Graph Neural Networks
Core Problem: Classical graphs and hypergraphs are insufficient to model highly complex, hierarchical, multi-level structures with uncertainty and multi-aspect attributes (heterogeneous, partially inconsistent information) found in real-world networks.
Key Innovation: Develops the theoretical foundations of SuperHyperGraph Neural Networks (SHGNNs) and Plithogenic Graph Neural Networks, extending message-passing principles to these advanced higher-order structures to model hierarchical, multi-level, and uncertain relational data.
375. ARCANE -- Early Detection of Interplanetary Coronal Mass Ejections
Core Problem: Robust real-time automatic detection of Interplanetary Coronal Mass Ejections (ICMEs) in streaming solar wind data for early warning systems remains a significant challenge under realistic operational constraints.
Key Innovation: Presents ARCANE, the first framework explicitly designed for early ICME detection in streaming solar wind data under operational constraints, utilizing a ResUNet++ model that significantly outperforms a threshold-based baseline in detecting high-impact events with minimal performance degradation from real-time data.
376. Domain Generalization and Adaptation in Intensive Care with Anchor Regression
Core Problem: Predictive models in clinical settings (e.g., ICU) suffer performance degradation when deployed in new hospitals due to distribution shifts, making them unreliable across heterogeneous multi-center data.
Key Innovation: Application and extension of anchor regression (with anchor boosting) for causality-inspired domain generalization on large-scale ICU data, demonstrating improved out-of-distribution performance and proposing a framework to quantify the utility of external datasets across different data regimes.
377. Distributional Shrinkage I: Universal Denoiser Beyond Tweedie's Formula
Core Problem: Denoising from noisy measurements where only the noise level is known, not the noise distribution, particularly when the goal is to recover the entire underlying signal distribution rather than individual realizations.
Key Innovation: Universal denoisers that are agnostic to a wide range of signal and noise distributions, offering order-of-magnitude improvements over Bayes-optimal denoisers (like Tweedie's formula) for recovering the entire signal distribution, achieving higher-order accuracy in matching generalized moments and density functions.
378. V-MORALS: Visual Morse Graph-Aided Estimation of Regions of Attraction in a Learned Latent Space
Core Problem: Existing reachability and safety analysis methods in robotics often require known system dynamics, large datasets, or full state information, making them impractical when only high-level sensor measurements, such as images, are available.
Key Innovation: Presents V-MORALS, a method that learns a latent space for reachability analysis from image-based trajectories of a system, enabling the generation of Morse Graphs and computation of Regions of Attraction (ROAs) without relying on full state knowledge.
379. A dynamic model updating method for multi-scenario underwater target recognition in ocean engineering applications
Core Problem: Static models for underwater target recognition (UTR) are inadequate in dynamic multi-scenario ocean engineering applications due to issues like class imbalance, inter-class similarity, and catastrophic forgetting in class-incremental learning.
Key Innovation: A dynamic rebalancing method within a class-incremental UTR (CI-UTR) framework, integrating an Inter-Class Similarity-aware Memory Management (ICSMM) strategy and an Inter-Phase Knowledge Aggregation (IPKA) module, which significantly mitigates forgetting and improves incremental accuracy for multi-scenario UTR.
380. A decision making strategy for cost-effective anchoring and mooring design in floating offshore wind systems
Core Problem: Optimizing stationkeeping systems for floating offshore wind turbines is essential for cost reduction and power production, but anchor system selection and design are often overlooked, lacking a cost-effective, data-minimal decision-making strategy.
Key Innovation: A flexible decision-making strategy for selecting the most cost-effective anchor at early design stages, based on minimal input (seabed type, anchor load angle, mooring properties) and a bottom-up cost estimation. This strategy provides consistent size and cost estimates and unifies fragmented information, enabling more realistic anchor system selection.
381. Few-shot recognition of slowly moving and small underwater targets with transfer learning and physics-informed fusion
Core Problem: Challenging few-shot recognition of slowly moving and small underwater targets due to their weak features, data scarcity, and poor sonar resolution.
Key Innovation: Proposing FUSTR-Net, a physics-informed transfer learning method that integrates multi-domain features (time, frequency, time-frequency) with azimuth-dependent scattering information as a latent prior, achieving high recognition accuracy even with limited training samples.
382. GLFFuse: A Multimodal Feature-Level Fusion Network for Multitask Fine-Grained Recognition of Arctic Sea Ice
Core Problem: Accurate monitoring and recognition of Arctic sea ice are essential, but existing studies lack sufficient exploration of jointly capturing high-frequency spatial details and low-frequency global structures from multisource observations, and effective information interaction across different data modalities.
Key Innovation: Proposes GLFFuse, a novel multimodal feature-level fusion network for fine-grained Arctic sea ice recognition. It integrates SAR, AMSR2, ERA5, and auxiliary data, combining a long short-range attention mechanism with an invertible neural network to jointly model global contextual patterns and local structural details, enhancing complementarity among multimodal features and improving prediction accuracy and stability across seasons.
383. Multilevel Alignment-Based Domain Generalization Network for Cross-Scene Hyperspectral Image Classification
Core Problem: Deep learning models struggle with generalization in cross-scene hyperspectral image (HSI) classification due to domain shift and inadequate alignment in existing domain generalization (DG) methods.
Key Innovation: Proposes a multilevel alignment framework for HSI DG, integrating data-level (residual-separated augmentation alignment), semantic-level (cross-modal semantic alignment with comprehensive text features), and knowledge-level (adaptive weight fusion with foundation models) mechanisms to mitigate domain shift and improve cross-scene classification.
384. CCMANet: A Cross-Layer Cascade Network With Multiattention Mechanisms for Remote Sensing Object Detection
Core Problem: Remote sensing object detection is challenging due to wide coverage, significant object scale variations, dense object distribution, and severe background interference in remote sensing images.
Key Innovation: Proposes CCMANet, a cross-layer cascade network with multiattention mechanisms, which incorporates a multiattention collaborative module for feature filtering, a maximum feature fusion module for diversity, and an improved dual-spatial pyramid pooling module for enriching target features, achieving higher detection accuracy.
385. Synergistic impact of simultaneously assimilating radar- and radiometer-based soil moisture retrievals on the performance of numerical weather prediction systems
Core Problem: Improving the accuracy of global soil moisture analysis and numerical weather prediction (NWP) performance by effectively integrating complementary soil moisture data from multiple satellite sensors (radar and radiometer).
Key Innovation: Evaluation of the synergistic impact of simultaneously assimilating ASCAT and SMAP soil moisture retrievals into the Korean Integrated Model (KIM) via a weakly coupled data assimilation framework, demonstrating enhanced global soil moisture analysis accuracy (up to 10.5%) and more balanced skill improvements for specific humidity and air temperature forecasts compared to single-sensor assimilation.
386. Investigating the influence of Fouta Jallon topography on the West African mean surface climate using RegCM5
Core Problem: The specific impact of the Fouta Djallon topography on the mean surface climate of West Africa, particularly during the June–September (JJAS) season, needs further investigation.
Key Innovation: Used the RegCM5 regional climate model to compare current topography with a flattened terrain scenario, demonstrating that reduced elevation leads to decreased rainfall and increased temperatures over the Guinea Coast, primarily due to suppressed orographic uplift and a northward shift of the ITCZ.
387. Extended Dual Porosity/Dual Permeability Model for Fluid Flow in Reservoirs with Multiscale Complex Fracture Networks
Core Problem: Accurately modeling fluid flow through naturally fractured reservoirs is computationally expensive and complex with traditional discrete fracture models (DFM), especially for large-scale, highly fractured systems.
Key Innovation: Introduced an extended dual porosity/dual permeability (xDPDP) formulation for single-phase flow in fractured reservoirs. This method implicitly represents fractures, combining homogenized permeability with a modified shape factor, allowing for regular, nonconforming meshes and significantly reducing meshing complexity while preserving the influence of multiple fracture types.
388. Correction: Laboratory Characterization and Modeling of Creeping Shale as Well Barrier
Core Problem: This item is a journal correction notice that updates previously published material rather than presenting new primary geohazard evidence.
Key Innovation: Improves record accuracy and reproducibility; it does not add new experimental, observational, or modeling results.
389. Relationships between sediment connectivity and hypsometry in an intertidal microtidal estuary
Core Problem: Understanding and quantifying functional surface sediment connectivity and transport pathways in shallow estuarine systems with multiple sub-basins is complex due to non-linear interactions between hydrodynamics and local morphology.
Key Innovation: Applies graph theory to depth-averaged numerical model outputs to quantify functional surface sediment connectivity between geomorphic regions in an estuary, revealing how patterns vary with hydrodynamic energy, grainsize, and morphology, and identifying source-sink relationships for improved predictive understanding of sediment exchange.
390. Mechanical control of invasive vegetation in China promotes tidal channels development: A case study of the Yellow River mouth wetland
Core Problem: The geomorphic response of tidal channel networks to large-scale mechanical control programs for invasive vegetation in coastal wetlands, and the underlying mechanisms, remain unclear.
Key Innovation: Integrates multi-temporal Sentinel-2 imagery and field surveys to quantify the spatiotemporal evolution of tidal channel networks in the Yellow River mouth wetland, revealing that mechanical control shifts networks from dynamic equilibrium to rapid development by creating depressions, removing vegetation, and altering sediment properties, offering a framework for predicting geomorphic responses to invasive species management.
391. Identification of marine algal blooms by SDGSAT-1 multispectral imagery
Core Problem: There is an emergent need for high-resolution satellite remote sensing to accurately detect and differentiate various types of marine algal blooms to support sustainable ocean and coastal zone development.
Key Innovation: This study explored the use of SDGSAT-1 Multispectral Imager (MII) for detecting and classifying five types of algal blooms, developing a spectrally interpretable decision tree model that outperformed traditional spectral indices and machine learning models, demonstrating the potential of SDGSAT-1 MII for high-resolution multispectral imagery-based monitoring of marine algal blooms.
392. Towards automated laser triangulation with stereo and planar constraints in refractive environments
Core Problem: Automated correspondence search for stereo laser triangulation in refractive environments (e.g., underwater) is challenging, requiring high precision and efficiency without iterative back-projection for applications like conservation monitoring.
Key Innovation: This paper presents two automated correspondence-search algorithms for stereo laser triangulation in refractive environments, enforcing coplanarity through forward ray tracing, which directly minimizes ray skew distance or intersection distance, achieving sub-millimeter precision and efficient reconstruction of timber structures in a water-filled tank, with improved noise reduction and fewer outliers.
393. Interaction between segmented lining and surrounding rock considering sliding layer for compressed air energy storage caverns
Core Problem: Monolithic linings in high-pressure underground energy storage caverns are prone to cracking and tensile damage due to significant shear stress, compromising structural stability and sealing integrity.
Key Innovation: Proposed a composite segmented lining system with a sliding layer to reduce shear stiffness and enhance deformation release. Numerical analyses and mechanical tests showed that reducing interfacial shear stiffness is more effective than increasing preset seams for alleviating lining stresses and maintaining structural integrity.
394. Experimental study on the effects of raindrop impact on particle transport across urban road surface
Core Problem: The coupled mechanisms and relative contributions of raindrop impact and runoff shear on particle transport across urban road surfaces, which is a primary source of stormwater pollution, remain unclear.
Key Innovation: Decoupled the effects of raindrop impact on particle transport under varying conditions through laboratory experiments. Demonstrated that raindrop impact significantly reduced runoff initiation threshold, enhanced initial particle detachment, and profoundly improved the final wash-off fraction of coarse particles, providing a mechanistic understanding for optimizing physically-based models of urban particle transport.
395. Uncertainty-aware reservoir operation projections using multi-model weighting and adaptive hedging rules
Core Problem: Conventional reservoir operation rules are vulnerable to climate change-driven inflow variability and amplify projection uncertainties, leading to future shortages.
Key Innovation: An integrative, uncertainty-aware framework combining a dynamic weighting scheme (UO-MME) for credible inflow projections with a two-dimensional hedging model (HDG-2d) to optimize reservoir releases, reducing uncertainty in reservoir outputs and providing a more robust basis for planning.
396. Transient response of deep-buried twin tunnels with arbitrary cross-sections under P- and SV-waves
Core Problem: Understanding the mutual interactions and transient response of deep-buried twin tunnels with arbitrary cross-sections under aperiodic P- and SV-wave disturbances is crucial for design but challenging to model theoretically and numerically.
Key Innovation: Develops a theoretical and numerical methodology combining modified conformal mapping, complex variable theory, multi-polar coordinates, and Fourier synthesis to obtain the transient response of twin arbitrary cross-sectional tunnels to P- and SV-waves, providing insights for reinforcement design.
397. Risk-based cumulative damage assessment of multi-span continuous bridges under mainshock-aftershock sequences: A novel state-dependent fragility surface framework
Core Problem: Current methods for assessing earthquake damage to structures rely on single-event threshold evaluation, failing to capture damage accumulation across mainshock-aftershock sequences and potentially underestimating structural vulnerability.
Key Innovation: Proposes a novel state-dependent fragility surface framework that couples conditional damage state transition paths with cumulative-damage fragility surfaces over a 2D IM space, effectively quantifying cumulative seismic damage and providing more reliable exceedance probabilities for bridges under MS-AS sequences.
398. Full-scale test on shield tunnel circumferential joints with oblique bolts and tongue-groove structures: Coupled shear-resisting behaviors
Core Problem: Shield tunnels experience non-uniform loading from external disturbances, leading to circumferential joint dislocations, and the dynamic coupled behavior between adjacent joints with oblique bolts and tongue-groove structures is not well understood.
Key Innovation: Full-scale three-ring testing and refined finite element analyses revealed asymmetric failure evolution, dynamic stiffness redistribution, and progressive load transfer between adjacent joints, quantifying the distinct behavioral patterns of compressed vs. tensioned bolt joints.
399. Vacuum dewatering behavior of foam-conditioned clay soil: Implications for foam optimization in earth pressure balance shield tunneling
Core Problem: Optimizing foam conditioning for clay soils in earth pressure balance (EPB) shield tunneling is crucial for efficient conditioning and recycling, but a simple and effective method to assess foam optimization based on dewatering behavior is lacking.
Key Innovation: Investigated the vacuum dewatering behavior of foam-conditioned clay soils, demonstrating a linear decrease in filtrate loss (dewaterability) with increasing mechanical indices (undrained shear strength, tangential/normal adhesion stress), providing a novel approach for evaluating foam conditioning effectiveness in EPB shield tunneling.