TerraMosaic Daily Digest: Feb 9, 2026
Daily Summary
This digest synthesizes 281 selected papers and focuses on landslide process mechanics and slope evolution, high-resolution remote-sensing monitoring workflows, infrastructure-focused hazard performance. Top-ranked studies examine satellite and LiDAR-based deformation monitoring, flood generation and hydroclimatic forcing, and operational early-warning thresholding.
Across the full set, evidence converges on mechanism-constrained analysis with operational relevance, especially for coastal and submarine hydro-geomechanics and seismic source-to-ground response pathways. The strongest contributions pair interpretable process evidence with monitoring or forecasting workflows that support warning design and risk prioritization.
Key Trends
- Landslide studies increasingly resolve process chains: Contributions connect triggering conditions, slope deformation, and mobility outcomes, improving the basis for warning thresholds and scenario testing.
- Monitoring workflows rely on integrated remote-sensing products: Multi-source satellite and airborne observations are used for deformation retrieval, change detection, and rapid post-event mapping.
- Infrastructure-facing outputs are increasingly decision-ready: Asset performance is evaluated with uncertainty-aware frameworks to support mitigation and maintenance prioritization.
- Coastal and submarine hazards are treated as coupled systems: Wave, mass-transport, and shoreline processes are analyzed together with engineering implications.
- Seismic hazard research links source behavior to ground response: Recurring topics connect rupture or loading conditions with geotechnical performance and consequence assessment.
Selected Papers
This digest features 281 selected papers from 1566 RSS items analyzed (out of 3264 raw RSS items scanned; 1569 new papers after deduplication). Each paper has been evaluated for its relevance to landslide and broader geohazard research and includes links to the original publications.
1. Analysis of triggering factors behind the October 2023 South Lhonak GLOF event in the Sikkim Himalaya using multiple remote sensing data
Core Problem: Urgent need to understand the triggering factors behind the catastrophic Glacial Lake Outburst Flood (GLOF) that occurred from the South Lhonak Glacial Lake in the Sikkim Himalaya in October 2023.
Key Innovation: Analysis of the triggering factors for the October 2023 South Lhonak GLOF event using multiple remote sensing data to provide insights into the causes of this catastrophic geohazard.
2. Role of sediment entrainment in the flash-flood to debris-flow transition during cascading landslide dam failures
Core Problem: Lack of understanding of the mechanisms by which rapid floods transform into dense debris flows during cascading landslide dam failures, particularly the role of sediment entrainment and its amplification effects.
Key Innovation: Experimental and numerical demonstration that successive landslide dam breaches and sediment entrainment create a positive feedback loop, significantly amplifying hydrodynamic parameters and transforming flash floods into destructive debris flows, even with low upstream reservoir volumes.
3. Bridging data gaps in landslide early warning: Physics-derived rainfall thresholds under rainfall uncertainty
Core Problem: The inability to establish reliable landslide early warning thresholds in data-scarce mountainous regions due to the absence of precise rainfall-landslide event pairing data, which cripples traditional empirical methods.
Key Innovation: A novel physics-derived framework that integrates the TRIGRS model with Frank Copula-GPD probabilistic rainfall modeling to define robust, probability-informed, four-tiered rainfall thresholds for landslide early warning, minimizing reliance on historical event pairing and demonstrating high spatiotemporal transferability.
4. Effects of grain size on landslide–forest interaction
Core Problem: Insufficient understanding of how forests mitigate landslide runout, specifically how grain size influences retention and jamming mechanisms, leading to oversimplified representations in existing models.
Key Innovation: Experimental investigation using reduced-scale flume tests that reveals distinct deposition and jamming mechanisms (frontal deposit-induced and arching-induced) for fine-grained and coarse-grained landslide flows interacting with forests, providing phase diagrams for determining minimum tree density for effective jamming.
5. A near-real-time multi-temporal polarimetric InSAR method for landslides monitoring in rapid-decorrelation scenarios
Core Problem: Traditional InSAR methods suffer from severe degradation in reliability and accuracy in regions with dense vegetation or frequent surface changes due to rapid temporal decorrelation, limiting their effectiveness for near-real-time landslide monitoring.
Key Innovation: Proposes a near-real-time sequential multi-temporal polarimetric InSAR (MT-PolInSAR) method that dynamically reselects statistically homogeneous pixels, applies sequential polarimetric-temporal phase optimization within short, high-coherence windows, and updates deformation time series through sequential least squares inversion, significantly increasing coherent pixel density and improving deformation accuracy in rapid-decorrelation areas.
6. An improved depth-averaged landslide dam breach model with modified sediment transport and bank collapse algorithms
Core Problem: Existing depth-averaged models for landslide dam breach simulation struggle to accurately represent longitudinal erosion and lateral bank collapse due to a lack of precise physical representations.
Key Innovation: Proposed an improved depth-averaged landslide dam breach model by integrating enhanced sediment transport (dynamic critical Shields parameter) and bank collapse algorithms (new algorithm using true 3D slope angle), validated against experiments and applied to the Tangjiashan landslide dam breach, achieving high agreement with measurements.
7. Wavelet Packet-Based Diffusion Model for Ground Motion Generation with Multi-Conditional Energy and Spectral Matching
Core Problem: Generating realistic ground motions that accurately match both temporal energy evolution (which affects structural response and damage) and target response spectra, a challenge for existing synthesis methods.
Key Innovation: Proposes a multi-conditional diffusion framework for ground motion synthesis using wavelet packet decomposition for signal representation. It integrates heterogeneous conditions (spectral ordinates, Arias intensity, temporal parameters, Husid curves) via a Transformer-based encoder, and explicitly constrains temporal energy, leading to improved control of energy onset and duration while preserving spectrum matching and diversity sampling.
8. AI-Driven Predictive Modelling for Groundwater Salinization in Israel
Core Problem: The increasing salinity and contamination of groundwater is a serious issue, and there is a need for a comprehensive understanding of its underlying causal factors and robust predictive models to identify important meteorological, geological, and anthropogenic drivers.
Key Innovation: Develops an AI-driven predictive modeling framework integrating various ML models (RF, XGBoost, NN, LSTM, CNN, LR) with RFE, GSA, XAI (SHAP), and causality analysis to comprehensively identify and understand key meteorological, geological, and anthropogenic drivers of groundwater salinization at a country scale, particularly highlighting the role of Treated Wastewater.
9. Automated rock joint trace mapping using a supervised learning model trained on synthetic data generated by parametric modelling
Core Problem: Automated rock joint trace mapping from images is challenging due to limited real data and class imbalance, hindering the training of supervised learning models.
Key Innovation: A geology-driven machine learning method that combines discrete fracture network models for synthetic data generation with supervised image segmentation, demonstrating that synthetic data can effectively support joint trace detection, especially when fine-tuned with a small amount of real data, leading to more geologically meaningful results.
10. MMLSv2: A Multimodal Dataset for Martian Landslide Detection in Remote Sensing Imagery
Core Problem: The need for a comprehensive, multimodal dataset to advance landslide detection and segmentation research, particularly for evaluating model robustness and generalization in challenging and geographically diverse regions.
Key Innovation: The MMLSv2 dataset, a multimodal dataset for Martian landslide segmentation comprising 664 images with seven bands (RGB, DEM, slope, thermal inertia, grayscale) and an isolated test set for evaluating spatial generalization, supporting stable training while highlighting challenges in complex landslide regions.
11. RAPTOR-AI for Disaster OODA Loop: Hierarchical Multimodal RAG with Experience-Driven Agentic Decision-Making
Core Problem: Humanitarian Assistance and Disaster Relief (HADR) operations require rapid synthesis of fragmented, multimodal information for time-critical decision-making under extreme uncertainty, which traditional information systems struggle to provide.
Key Innovation: Introduces RAPTOR-AI, an agentic multimodal Retrieval-Augmented Generation (RAG) framework that provides dynamic, experience-driven decision support for disaster response through hierarchical multimodal knowledge construction, entropy-aware agentic control, and experiential knowledge integration, demonstrating significant improvements in retrieval precision, situational grounding, and task decomposition accuracy.
12. Impact of sea level rise on storm surge dynamics during cold surges in the northern East China Sea: Relevance of surge variability to semidiurnal tidal regime
Core Problem: The impact of sea level rise (SLR) on storm surge dynamics during cold surge events, particularly how it modifies surge magnitude, timing, spectral structure, and tidal interactions, is not fully understood, hindering accurate coastal hazard assessments.
Key Innovation: Conducted numerical simulations for historical cold surge events under multiple SLR scenarios, revealing that SLR induces spatially heterogeneous and nonlinear responses in storm surge behavior, generally decreasing surge amplitudes but increasing maximum storm tide, and highlighted the critical role of high-frequency surge components and nonlinear tidal interactions for coastal hazard assessments.
13. The efficacy of green and gray coastal structures in storm surge mitigation
Core Problem: Limited quantitative comparison of wave-attenuation efficacy between nature-based solutions (mangroves) and gray coastal structures (submerged breakwaters) under storm surge conditions.
Key Innovation: Full-scale numerical simulations (OpenFOAM) to compare wave-attenuation efficacy of mangroves and submerged breakwaters, proposing new empirical formulations for wave transmission coefficients. Findings provide practical insights for hybrid coastal defense strategies.
14. Innovative polygonal coastal index concept and probabilistic vulnerability predictions
Core Problem: Limitations of the classical coastal vulnerability index (CVI) which uses five mutually exclusive integer scores, potentially oversimplifying complex coastal vulnerability assessments.
Key Innovation: Proposes an innovative polygonal coastal index concept and probabilistic vulnerability predictions by plotting US-ranked CVI input variables in a polygonal graph, aiming to provide a more nuanced and comprehensive assessment of coastal vulnerability.
15. Chronological development of gravitational slope deformation induced by upstream knickpoint migration
Core Problem: Understanding the structural causes and chronological development of deep-seated gravitational slope deformation (DGSD) in a specific geological context, particularly its long-term evolution and influencing factors.
Key Innovation: Investigated DGSD in the Chichibu area using field surveys, high-resolution DEM, drilling, and tephrochronological dating. Identified low-angle thrust faults exhumed by river erosion as the cause, dating the deformation over 200,000 years with typical displacement rates and noting recent acceleration and potential deceleration during glacial ages.
16. Dynamic penetration test-based probabilistic hyperbolic model for evaluating liquefaction potential of gravelly soils
Core Problem: The challenge of accurately and practically evaluating liquefaction potential in gravelly soils, which are susceptible to seismic liquefaction, due to the limitations of conventional in-situ testing methods.
Key Innovation: Development of accurate, straightforward, and depth-consistent probabilistic and deterministic hyperbolic models for evaluating liquefaction potential in gravelly soils, based on dynamic penetration test (DPT) data and a global database, suitable for seismic hazard assessment.
17. Flow behavior and rheology of rock-ice granular mixtures with pendular liquid bridges
Core Problem: The flow mechanisms of rock-ice avalanches, which pose increasing threats in cold mountainous regions, remain poorly understood.
Key Innovation: Numerical simulations reveal that pendular liquid bridges, even if weak, enhance the integrity of rock-ice granular mixtures and impede segregation, influencing their rheology and motion, with μ(I) rheologies appearing more applicable than Voellmy for wet granular flows.
18. Bridging factor of safety prediction and strength reduction workflows using an interpretable Transformer-enhanced ensemble model
Core Problem: Improving the computational efficiency of slope stability analysis using the strength reduction method (SRM) while maintaining physical interpretability and effectively integrating data-driven FoS prediction into established workflows.
Key Innovation: Proposed an SRM-Transformer stacked ensemble model (TSEM) framework that integrates ML-based FoS prediction with physics-based SRM, achieving high accuracy (R2 = 0.9857), reducing computational time by over 47%, and decreasing strength reduction steps by nearly 58% while maintaining physical consistency.
19. 2D Broadband Magnetotelluric Study of the Axial Fault Region of the New Madrid Seismic Zone
Core Problem: Determining whether seismicity in the New Madrid Seismic Zone (NMSZ), specifically the Axial Fault, is predominantly influenced by fluids or fault locking, as previous resistivity studies had limited resolution.
Key Innovation: Conducted a comprehensive broadband magnetotelluric (MT) study along five profiles, revealing that earthquakes cluster within a highly resistive zone (1,000–10,000 Ω-m), suggesting a fault-locking mechanism. Also identified a conductive anomaly on the SE side of the fault, interpreted as a weaker, more ductile zone, suggesting both fault locking and fluid presence influence NMSZ seismicity.
20. Lagged backward-compatible physics-informed neural networks for unsaturated soil consolidation analysis
Core Problem: Simulating and inverting one-dimensional unsaturated soil consolidation under long-term loading is challenging due to coupled air and water pressure dissipation across multi-scale time domains.
Key Innovation: Develops a Lagged Backward-Compatible Physics-Informed Neural Network (LBC-PINN) that integrates logarithmic time segmentation, lagged compatibility loss enforcement, and segment-wise transfer learning to accurately predict pore air and pore water pressure evolution in unsaturated soil consolidation, validated against FEM results.
21. Contactless estimation of continuum displacement and mechanical compressibility from image series using a deep learning based framework
Core Problem: Conventional methods for contactless estimation of continuum displacement and mechanical compressibility from optical observations are time-consuming and not suitable for high-throughput data processing.
Key Innovation: An efficient deep learning based end-to-end framework using two deep neural networks for image registration and material compressibility estimation, which accurately determines material compressibility and displacement from image series, outperforming conventional approaches in efficiency and accuracy.
22. Perspective-aware fusion of incomplete depth maps and surface normals for accurate 3D reconstruction
Core Problem: Existing orthographic gradient-based depth-normals fusion methods for 3D reconstruction from depth and surface normal maps do not explicitly account for perspective projection, leading to metrically inaccurate reconstructions, and struggle with missing depth measurements.
Key Innovation: Proposes a perspective-aware log-depth fusion approach that extends existing methods by explicitly accounting for perspective projection for metrically accurate 3D reconstructions and leverages available surface normal information to inpaint gaps in depth measurements.
23. Building Damage Detection using Satellite Images and Patch-Based Transformer Methods
Core Problem: Rapid and accurate building damage assessment post-disaster using satellite imagery is challenging due to label noise and severe class imbalance in existing datasets, hindering the performance of damage classification models.
Key Innovation: A targeted patch-based pre-processing pipeline combined with a frozen-head fine-tuning strategy for small Vision Transformer (ViT) architectures, which achieves competitive multi-class damage classification performance by effectively isolating structural features and minimizing background noise.
24. A Machine Learning accelerated geophysical fluid solver
Core Problem: Applying machine learning to solve partial differential equations (PDEs) in geophysical fluid dynamics while maintaining mathematical constraints and improving the accuracy and stability of low-resolution simulations.
Key Innovation: Proposes and implements data-driven discretization methods using deep neural networks to accelerate and improve geophysical fluid solvers (e.g., shallow water and Euler equations), demonstrating improved accuracy and stability compared to traditional schemes.
25. Chamelion: Reliable Change Detection for Long-Term LiDAR Mapping in Transient Environments
Core Problem: Online change detection and long-term map maintenance for mobile robots are challenging in dynamic, transient environments (e.g., construction sites) due to frequent occlusions and spatiotemporal variations, and lack of real-world annotated data.
Key Innovation: Proposes Chamelion, a dual-head network for reliable online change detection and long-term LiDAR map maintenance, coupled with a novel data augmentation strategy that synthesizes structural changes, enabling effective model training without extensive ground-truth annotations and demonstrating strong generalization across diverse scenarios.
26. Neural MJD: Neural Non-Stationary Merton Jump Diffusion for Time Series Prediction
Core Problem: Deep learning methods for time series prediction often lack interpretability and struggle to generalize to non-stationary data, particularly when abrupt changes (jumps) occur, as they don't explicitly model underlying stochastic processes.
Key Innovation: Introduces Neural MJD, a neural network-based non-stationary Merton jump diffusion model that explicitly formulates forecasting as an SDE simulation, combining Itô diffusion for non-stationary dynamics and a compound Poisson process for abrupt jumps. It enables tractable learning via likelihood truncation and proposes an Euler-Maruyama with restart solver, outperforming state-of-the-art methods.
27. XiChen: A global weather observation-to-forecast machine learning system via four-dimensional variational gradient-guided flexible assimilation
Core Problem: Most ML weather forecasting systems rely on initial conditions from Numerical Weather Prediction (NWP) and lack operational robustness when observation sources change, requiring redesign or retraining.
Key Innovation: XiChen, a global weather observation-to-forecast ML system that uses four-dimensional variational (4DVar) gradient-guided flexible assimilation to map heterogeneous observations into a common state space, achieving competitive forecasting metrics with operational NWP systems and providing a physically consistent route for ML-based global weather forecasting.
28. China Regional 3km Downscaling Based on Residual Corrective Diffusion Model
Core Problem: Efficiently producing high-resolution numerical weather forecasts, particularly for large regions and multiple atmospheric variables, from lower-resolution global model outputs.
Key Innovation: Applying and enhancing a diffusion-based downscaling framework (CorrDiff) with a global residual connection to generate 3km weather forecasts for the China region from 25km global models, demonstrating superior accuracy compared to a regional operational model (CMA-MESO) and generating realistic fine-scale details.
29. Experimental study on soil erosion induced by leakage of circumferential joints in shield tunnels
Core Problem: Leakage at shield tunnel joints is a primary cause of soil erosion and ground collapse, and the mechanisms and influencing factors (like leakage location and presence of other tunnels) are not fully understood.
Key Innovation: Conducted model tests to reveal a three-stage evolution of soil erosion (internal erosion, surface collapse, stabilization), quantified the extent of ground collapse for different leakage locations, and demonstrated that existing upper tunnels significantly mitigate soil erosion and collapse width.
30. Statistical evaluation of joint occurrence among wind, wave, and seismic loads for offshore wind turbine design
Core Problem: Current design standards for offshore wind turbines may not accurately reflect the joint occurrence frequencies of wind, wave, and seismic loads, potentially leading to overly conservative designs or overlooking critical multi-hazard scenarios.
Key Innovation: Constructed a unique dataset combining long-term earthquake observations with concurrent wind-wave conditions, statistically evaluated the joint occurrence characteristics of these three loads, and identified specific critical scenarios (e.g., high swell under weak wind coinciding with seismic motion) to refine multi-hazard offshore wind turbine design.
31. The impact of onshore wind on regular wave overtopping at Accropode-armoured seawall
Core Problem: The influence of onshore wind on wave overtopping discharge at seawalls is insufficiently quantified, leading to inaccuracies in overtopping prediction and coastal engineering design.
Key Innovation: Combined physical experiments with a validated wind-wave coupled numerical model to systematically investigate the impact of onshore wind on wave overtopping at Accropode-armoured seawalls, demonstrating that wind significantly increases overtopping discharge, and proposed an empirical correction formula to improve prediction accuracy.
32. Emergence of climate change signal in CMIP6 extreme indices
Core Problem: The increasing frequency of climate and weather extremes due to anthropogenic climate change necessitates understanding when and where these changes will occur (time of emergence) to inform mitigation and adaptation measures.
Key Innovation: A comprehensive investigation of the time of emergence for 29 climate extreme indices using a weighted ensemble of 21 CMIP6 models, revealing distinct spatial and temporal emergence patterns for different temperature and precipitation indices, and highlighting regions of early emergence and model disagreement.
33. GloWE-8D: a global long-term 8-day wind erosion dataset from 1982 to 2020
Core Problem: Lack of a long-term, high spatiotemporal resolution, and publicly available global-scale wind erosion dataset, which has constrained a deeper understanding of its dynamic processes and driving mechanisms.
Key Innovation: Construction of GloWE-8D, a global long-term (1982-2020) 8-day wind erosion dataset with 0.05° spatial resolution, enhanced by a residue factor scheme for improved characterization of wind erosion suppression during vegetation cover periods, providing crucial data for dust emission estimation and land degradation prevention.
34. Research on electrical structure detection of surface subsidence based on ground-airborne frequency-domain electromagnetic method
Core Problem: Rapidly and effectively detecting subsurface electrical anomalies related to surface subsidence, especially in complex terrains, and providing geophysical evidence for interpreting the causes of subsidence.
Key Innovation: Evaluated the ground-airborne frequency-domain electromagnetic method (GAFEM) for detecting subsurface electrical anomalies in a surface subsidence area. Found that low-resistivity anomalies spatially correlate with severe subsidence, and the probability of subsidence in low-resistivity zones is significantly higher, demonstrating its potential for identifying potential surface subsidence hazards.
35. Self-enhancing climatic resilience of surface soil through bio-carbonation constructed barrier
Core Problem: The amplified risk of geological hazards like slope failure and earthen infrastructure instability due to intensified matter and energy exchanges between atmosphere and surface soil under extreme climatic events.
Key Innovation: Proposal and systematic investigation of a bio-carbonation constructed barrier that demonstrates self-enhancing climatic resilience of surface soil through atmospheric CO2-driven carbonation and crystal reorganization, providing a sustainable strategy for geohazard prevention and mitigation.
36. Flood-Induced Traffic Congestion and Accessibility Loss for Urban Road Networks Using Agent-Based Simulation: The Case Study of Bristol, UK
Core Problem: Existing methods for assessing flood impact on urban road networks often don't fully capture dynamic traffic redistribution, congestion, and accessibility loss, especially considering agent-level behavioral responses and impacts on critical facilities.
Key Innovation: Proposes a methodology using agent-based traffic simulation (MATSim) to model dynamic traffic redistribution and congestion under flood conditions, comparing spatial shifts of congestion hotspots and integrating hazard scenarios to predict future congestion and accessibility impacts on critical facilities, providing a transferable framework for assessing urban transport resilience during flood events.
37. Steady-state and transient pressure conditions around the EPB-TBM: Numerical modelling based on experimental data
Core Problem: Accidental transient variations in the pressure boundary conditions applied by Earth Pressure Balance Tunnel Boring Machines (EPB-TBMs) can significantly increase soil settlements and impact pile behavior, leading to sudden displacements and negative skin friction, which are not fully understood or accurately predicted.
Key Innovation: Developed a three-dimensional numerical model of an EPB-TBM, calibrated and validated with experimental data, to simulate steady-state and transient pressure conditions, demonstrating that transient pressure phases can cause rapid increases in soil and pile displacement (e.g., 30140% of final settlement from a 35% frontal pressure reduction).
38. Evaluation of the effects of embedded depth and earthquake intensity on seismic response of monopile-supported offshore wind turbines through seismic centrifuge tests
Core Problem: Evaluating the seismic response and potential for permanent tilting of monopile-supported offshore wind turbines, particularly the effects of earthquake-induced excess pore water pressure and liquefaction, and the influence of embedded depth.
Key Innovation: Investigation of OWT seismic response using geotechnical centrifuge tests with scaled models, varying earthquake intensities and monopile slenderness ratios, demonstrating the impact of embedment depth and liquefaction on tilting and providing benchmarks for numerical models and design strategies for seismic resilience.
39. Probability integrated hybrid machine learning models for predicting surface settlement of PVD-treated soft soil
Core Problem: Critical challenge in predicting surface settlement of soft soil under PVD-embankment systems due to complex, time-dependent soil behavior, and the need to incorporate prediction uncertainty.
Key Innovation: Development of a probability-integrated hybrid machine learning model (CATB enhanced by ASRS) using in situ data to accurately predict surface settlement of PVD-treated soft soil, providing prediction confidence and actionable insights for optimized design.
40. Geospatial-Reasoning-Driven Vocabulary-Agnostic Remote Sensing Semantic Segmentation
Core Problem: Existing open-vocabulary semantic segmentation methods in remote sensing are "appearance-based," lacking geospatial contextual awareness, which leads to semantic ambiguity and misclassification for land-cover classes with similar spectral features but distinct semantic attributes.
Key Innovation: The Geospatial Reasoning Chain-of-Thought (GR-CoT) framework, which enhances MLLMs with geospatial contextual awareness through an offline knowledge distillation stream and an online instance reasoning stream, generating an image-adaptive vocabulary for precise pixel-level alignment with geographical semantics.
41. Thegra: Graph-based SLAM for Thermal Imagery
Core Problem: Thermal imagery, while practical for visual SLAM in visually degraded environments, often exhibits low texture, low contrast, and high noise, complicating feature-based SLAM and reliable mapping.
Key Innovation: Introduction of Thegra, a sparse monocular graph-based SLAM system for thermal imagery that leverages general-purpose learned features (SuperPoint, LightGlue) adapted with a preprocessing pipeline, and incorporates keypoint confidence scores into a confidence-weighted factor graph to improve estimation robustness without dataset-specific training.
42. AnomSeer: Reinforcing Multimodal LLMs to Reason for Time-Series Anomaly Detection
Core Problem: Multimodal Large Language Models (MLLMs) for Time-Series Anomaly Detection (TSAD) rely on coarse time-series heuristics and struggle with multi-dimensional, detailed reasoning, which is vital for understanding complex time-series data.
Key Innovation: AnomSeer, a framework that reinforces MLLMs to ground reasoning in precise, structural details of time series. It generates an expert chain-of-thought trace from classical analyses and uses a novel time-series grounded policy optimization (TimerPO) to unify anomaly classification, localization, and explanation.
43. On Geometry-Enhanced Parameter-Efficient Fine-Tuning for 3D Scene Segmentation
Core Problem: Adapting large-scale pre-trained point cloud models for 3D scene understanding to specific downstream tasks is computationally expensive (full fine-tuning), and existing parameter-efficient fine-tuning (PEFT) methods fail to adequately capture the geometric and spatial complexities of 3D point clouds.
Key Innovation: Introduces the Geometric Encoding Mixer (GEM), a novel geometry-aware PEFT module for 3D point cloud transformers, which explicitly integrates fine-grained local positional encodings with a lightweight latent attention mechanism. This allows for performance comparable to or exceeding full fine-tuning while updating only 1.6% of parameters, significantly reducing training time and memory.
44. Coupled seepage stress analysis and design optimization of deep buried subsea tunnels
Core Problem: Groundwater ingress and associated hydrostatic pressure pose critical challenges to the structural integrity of deep-buried underwater tunnels, with existing analytical solutions often oversimplifying hydraulic boundary conditions or neglecting seepage-induced body forces.
Key Innovation: Developed a closed-form analytical framework for coupled seepage-stress analysis of deep-buried circular tunnels under semi-infinite hydraulic boundary conditions, incorporating seepage pressure gradient as a volumetric load, and identifying an optimal permeability window for watertightness-oriented design and vulnerability assessment.
45. An experimental study of interaction process between sea ice and variable stiffness elastic plates at various speeds
Core Problem: Insufficient understanding of how structural deformation (due to variable stiffness) influences the interaction process between sea ice and structures, affecting ice failure modes and load magnitudes.
Key Innovation: An experimental study using small-scale indentation tests with elastic plates and frozen ice, demonstrating that structural deformation modifies relative velocity, high-pressure zone distribution, and ice failure modes, impacting load magnitude, and providing insights into the effect of structural stiffness and loading rates.
46. Experimental Investigation on Profile Evolution of Beach-Dune System Exposed to Irregular Waves
Core Problem: Understanding how beach-dune systems respond to irregular waves, particularly the mechanisms governing dune erosion and accretion, which is crucial for assessing coastal resilience and designing effective restoration strategies.
Key Innovation: Experimental investigation reveals that dune evolution (erosion vs. accretion) is governed by the relative elevation between total water level and dune crest (swash vs. overwash regime), and that while most sediment on the sand flat is transported seaward, the beach indirectly influences dune evolution by modulating wave energy at the dune toe.
47. A Hierarchical Optimization Model Based on Multisource Remote Sensing to Correct FABDEM
Core Problem: The Forest and Buildings Removed DEM (FABDEM) still exhibits substantial elevation errors in forested regions, limiting its accuracy for understory terrain extraction, which is crucial for various environmental and geohazard applications.
Key Innovation: A hierarchical optimization model integrating multisource remote sensing data and machine learning (Random Forest) to significantly enhance FABDEM accuracy, achieving up to 50% improvement in RMSE for understory terrain reconstruction, particularly in complex forested environments.
48. Dynamic response analysis of monopile-supported 10 MW offshore wind turbine under wind, wave and seismic loads
Core Problem: Lack of systematic comparative analyses of the dynamic behavior of monopile-supported offshore wind turbines under complex multi-hazard conditions (wind, wave, seismic loads) and varying soil parameters.
Key Innovation: Development of a numerical model based on the beam on nonlinear Winkler foundation method to investigate the effects of soil damping, added mass, scour depth, and soil strength variability on OWT fundamental frequency and dynamic responses under various combined wind, wave, and seismic loads, providing insights for performance-based design.
49. Research on the damping effect of inertial particle absorber on the seismic response of continuous rigid-frame bridge with super high piers
Core Problem: Lack of effective damping devices for complex bridge structures like continuous rigid-frame bridges with super high piers to improve their seismic performance and resilience.
Key Innovation: Proposal and experimental validation (shaking table tests) of an inertial particle absorber (IPA) for continuous rigid-frame bridges, demonstrating its effectiveness in suppressing high-order modes, increasing natural frequencies, and significantly reducing seismic response (crack width, acceleration, displacement, strain) with a small additional mass ratio.
50. Filter cake behaviors in sandy strata: From macro-parameters to pore structure
Core Problem: The critical role of slurry infiltration and filter cake formation in slurry shield tunneling, especially in sandy strata, and the need to understand the impact of pore structures on filter cake properties to ensure tunnel stability.
Key Innovation: Investigates pressure infiltration behaviors of bentonite slurry in sandy strata, identifying five types of filter cakes and establishing linear relationships between filter cake permeability and sand strata pore diameter, highlighting the controlling role of pore structure for tunneling stability.
51. Reduced Future Summer Water Availability in the Tien Shan Due To Glacier Wastage
Core Problem: Projecting future water availability in the Tien Shan mountains, a crucial source for over 100 million people, given the ongoing glacier wastage and its implications for water scarcity, especially during peak demand seasons.
Key Innovation: Dynamically modeled the evolution of all Tien Shan glaciers and merged their runoff with hydrological simulations from three global hydrological models to provide a hybrid estimate of future water availability. Findings project significant glacier mass loss, an initial increase in runoff followed by a substantial decrease, and a shift in peak runoff from summer to spring, significantly increasing the probability of unmet summer water demand.
52. Exploring Polarimetric Properties Preservation during Reconstruction of PolSAR images using Complex-valued Convolutional Neural Networks
Core Problem: The inherently complex-valued nature of Polarimetric SAR (PolSAR) data is often converted to real-valued representations for deep learning, potentially losing essential physical characteristics.
Key Innovation: Complex-valued Convolutional AutoEncoders are shown to effectively compress and reconstruct fully polarimetric SAR data while preserving crucial physical characteristics (e.g., Pauli, Krogager, Cameron, H-alpha decompositions), paving the way for robust, physics-informed complex-valued generative models for SAR data processing.
53. Risk-Sensitive Exponential Actor Critic
Core Problem: Current policy gradient methods for optimizing the entropic risk measure in deep reinforcement learning suffer from high-variance and numerically unstable updates, limiting their application to complex, real-world tasks requiring risk-aware agents.
Key Innovation: Provides theoretical justification for policy gradient methods on the entropic risk measure and proposes risk-sensitive exponential actor-critic (rsEAC), an off-policy model-free approach that avoids explicit representation of exponential value functions and their gradients, leading to more numerically stable updates and reliable learning of risk-sensitive policies.
54. Seeing Roads Through Words: A Language-Guided Framework for RGB-T Driving Scene Segmentation
Core Problem: Robust semantic segmentation of road scenes for autonomous driving remains challenging under adverse illumination, lighting, and shadow conditions, as existing RGB-Thermal fusion methods use static strategies that propagate noise.
Key Innovation: Proposes CLARITY, a language-guided framework that dynamically adapts its RGB-Thermal fusion strategy based on detected scene conditions using VLM priors. It also preserves dark-object semantics and uses a hierarchical decoder for structural consistency, achieving new state-of-the-art on the MFNet dataset.
55. Adaptive Image Zoom-in with Bounding Box Transformation for UAV Object Detection
Core Problem: Detecting small and sparse objects in UAV-captured images is challenging, hindering the optimization of effective object detectors.
Key Innovation: Introduces ZoomDet, an adaptive zoom-in framework with a lightweight offset prediction scheme, a novel box-based zooming objective, and a corner-aligned bounding box transformation method to significantly improve object detection on UAV images.
56. CA-YOLO: Cross Attention Empowered YOLO for Biomimetic Localization
Core Problem: Existing target localization systems often face limitations in both accuracy and the ability to recognize small targets in modern complex environments.
Key Innovation: Proposes CA-YOLO, a bionic stabilized localization system that integrates bionic modules (small target detection head, Characteristic Fusion Attention Mechanism) into the YOLO backbone and develops a bionic pan-tilt tracking control strategy, significantly enhancing accuracy and small target recognition.
57. TASTE: Task-Aware Out-of-Distribution Detection via Stein Operators
Core Problem: Existing out-of-distribution (OOD) detection methods are either data-centric (ignoring model impact) or model-centric (relying on classifier outputs without data geometry), failing to explicitly link distribution shift to the input sensitivity of the model or provide interpretable localization.
Key Innovation: TASTE (Task-Aware STEin operators), a task-aware framework based on Stein operators that links distribution shift to model input sensitivity. It provides theoretical guarantees, a clear geometric interpretation, and enables both detection and coordinate-wise localization of shifts, offering interpretable per-pixel diagnostics for image data.
58. Picasso: Holistic Scene Reconstruction with Physics-Constrained Sampling
Core Problem: Geometrically accurate 3D scene reconstructions often result in physically incorrect or implausible configurations (e.g., object interpenetration, unstable equilibrium) due to occlusions and measurement noise, hindering accurate dynamic behavior prediction for digital twins.
Key Innovation: Introduces Picasso, a physics-constrained reconstruction pipeline that holistically considers geometry, non-penetration, and physics for multi-object scenes, leveraging a fast rejection sampling method guided by an inferred object contact graph to produce physically plausible and geometrically accurate reconstructions. Also provides the Picasso dataset and a physical plausibility metric.
59. DAS-SK: An Adaptive Model Integrating Dual Atrous Separable and Selective Kernel CNN for Agriculture Semantic Segmentation
Core Problem: Semantic segmentation in high-resolution agricultural imagery requires models that are both accurate and computationally efficient for deployment on edge devices like UAVs, overcoming limitations of large dataset requirements and high computational costs.
Key Innovation: Proposes DAS-SK, a novel lightweight CNN architecture that integrates selective kernel convolution (SK-Conv) into dual atrous separable convolution (DAS-Conv) and enhances ASPP, achieving state-of-the-art performance with significantly fewer parameters and GFLOPs for efficient high-resolution remote sensing.
60. Efficient-SAM2: Accelerating SAM2 with Object-Aware Visual Encoding and Memory Retrieval
Core Problem: The Segment Anything Model 2 (SAM2) exhibits excellent performance in video object segmentation but suffers from a heavy computational burden, hindering its application in real-time video processing due to redundant computation in background regions and full-token memory attention.
Key Innovation: Efficient-SAM2, which accelerates SAM2 by leveraging object-aware sparse perception patterns, introducing object-aware Sparse Window Routing (SWR) for the image encoder to route background regions to a lightweight shortcut, and object-aware Sparse Memory Retrieval (SMR) for memory attention to only use salient memory tokens, achieving significant speedup with minimal accuracy drop.
61. The Connection between Kriging and Large Neural Networks
Core Problem: Despite the widespread impact of AI, the relationship between traditional spatial statistics models like Kriging and modern machine learning models, particularly neural networks, remains underexplored, hindering the development of more interpretable, reliable, and spatially aware ML techniques.
Key Innovation: Systematically explores and revisits the connections between Kriging (and Gaussian process regression) and large neural networks, highlighting their strong relationship and suggesting that combining both perspectives can enhance ML techniques by making them more interpretable, reliable, and spatially aware for various applications.
62. Understanding and Optimizing Attention-Based Sparse Matching for Diverse Local Features
Core Problem: Attention-based sparse image matching models, despite advances, have overlooked critical design choices impacting performance, and the roles of detectors versus descriptors in transformer-based matching are not fully understood, limiting their universality across diverse local features.
Key Innovation: Identifies a critical design choice impacting LightGlue performance, demonstrates that detectors are often the primary cause of performance differences, and proposes a novel approach to fine-tune existing models using keypoints from diverse detectors, resulting in a universal, detector-agnostic model that achieves or exceeds the accuracy of specifically trained models.
63. Low Rank Transformer for Multivariate Time Series Anomaly Detection and Localization
Core Problem: Existing Multivariate Time Series (MTS) anomaly diagnosis methods (detection and localization) offer limited theoretical insights, especially for anomaly localization, which is a vital but largely unexplored area for ensuring safety and reliability in complex systems.
Key Innovation: Proposes the Attention Low-Rank Transformer (ALoRa-T) model, applying low-rank regularization to self-attention and introducing the Attention Low-Rank score to capture temporal anomaly characteristics. Develops ALoRa-Loc for anomaly localization by quantifying interrelationships among time series, outperforming state-of-the-art methods.
64. Time-Delayed Transformers for Data-Driven Modeling of Low-Dimensional Dynamics
Core Problem: There is a need for data-driven models that can effectively capture unsteady spatio-temporal dynamics, bridging the gap between interpretable linear operator-based methods and powerful but complex deep sequence models, especially for nonlinear and chaotic regimes.
Key Innovation: Proposes the time-delayed transformer (TD-TF), a minimal transformer architecture interpreted as a nonlinear generalization of time-delayed dynamic mode decomposition (TD-DMD). It achieves linear computational complexity and small parameter count, matching linear baselines on near-linear systems and significantly outperforming them in nonlinear/chaotic regimes for long-term dynamics.
65. Automatic regularization parameter choice for tomography using a double model approach
Core Problem: Image reconstruction in X-ray tomography, especially with limited data, is an ill-posed inverse problem, and its effectiveness hinges on the difficult and often manual choice of a regularization parameter.
Key Innovation: A novel method for automatic regularization parameter selection based on using two distinct computational discretizations of the same problem, where a feedback control algorithm dynamically adjusts regularization strength to achieve sufficient similarity between reconstructions on the two grids.
66. Overview and Comparison of AVS Point Cloud Compression Standard
Core Problem: The large data size of point clouds poses significant challenges to their transmission and storage, hindering their widespread deployment in various applications, including those critical for 3D environmental mapping.
Key Innovation: This paper provides an overview and comparison of the AVS PCC standard, China's first-generation point cloud compression standard. It reviews its new coding tools and techniques, contrasting them with existing MPEG standards (G-PCC and V-PCC), to address the need for efficient point cloud data handling.
67. LEFT: Learnable Fusion of Tri-view Tokens for Unsupervised Time Series Anomaly Detection
Core Problem: Unsupervised time series anomaly detection (TSAD) is challenging because many anomalies are subtle and only manifest as inconsistencies across multiple data views (time, frequency, multi-resolution), and existing cross-view methods lack analysis-synthesis consistency.
Key Innovation: Presents LEFT (Learnable Fusion of Tri-view Tokens), a unified unsupervised TSAD framework that learns feature tokens from time, frequency, and multi-scale views. It introduces adaptive Nyquist-constrained spectral filters, a novel objective for reconstructing fine-grained targets from coarser structures, and an innovative time-frequency cycle consistency constraint to regularize cross-view agreement.
68. StretchTime: Adaptive Time Series Forecasting via Symplectic Attention
Core Problem: Transformer architectures for time series forecasting typically rely on positional encodings that assume uniform temporal progression, failing to represent 'time-warped' dynamics common in real-world systems.
Key Innovation: Formalizes the misalignment of time-warped dynamics and proves the mathematical incapability of RoPE. Proposes Symplectic Positional Embeddings (SyPE), a learnable encoding framework derived from Hamiltonian mechanics that generalizes RoPE, modulated by a novel input-dependent adaptive warp module. This mechanism, implemented in StretchTime, achieves state-of-the-art performance and superior robustness on datasets exhibiting non-stationary temporal dynamics.
69. Flow-Based Conformal Predictive Distributions
Core Problem: Conformal prediction sets are difficult to represent and use in high-dimensional or structured output spaces, limiting their integration with downstream tasks like sampling and probabilistic forecasting.
Key Innovation: Shows that any differentiable nonconformity score induces a deterministic flow, leading to a computationally efficient, training-free method for sampling conformal boundaries. This enables forming pointwise prediction sets and conformal predictive distributions, evaluated on applications like precipitation downscaling and hurricane trajectory forecasting.
70. Do physics-informed neural networks (PINNs) need to be deep? Shallow PINNs using the Levenberg-Marquardt algorithm
Core Problem: The need for efficient and accurate solutions for forward and inverse problems of nonlinear partial differential equations (PDEs), and questioning the necessity of deep architectures in Physics-Informed Neural Networks (PINNs).
Key Innovation: Demonstrating that shallow PINNs, when combined with the Levenberg-Marquardt algorithm for optimization, can provide accurate and computationally efficient solutions for a wide class of PDEs, outperforming BFGS.
71. DNS: Data-driven Nonlinear Smoother for Complex Model-free Process
Core Problem: Estimating hidden state sequences of complex, model-free dynamical processes from noisy, linear measurement sequences without prior knowledge of the nonlinear dynamics.
Key Innovation: Proposing a data-driven nonlinear smoother (DNS) with a recurrent architecture that provides a closed-form posterior of the hidden state sequence, learns in an unsupervised manner, and significantly outperforms existing smoothers for stochastic dynamical processes.
72. Empirically Understanding the Value of Prediction in Allocation
Core Problem: Institutions increasingly use prediction for scarce resource allocation, but it's unclear how to quantify the bottom-line welfare impact of investments in prediction versus other policy levers like expanding capacity or improving treatment quality.
Key Innovation: Develops an empirical toolkit ('rvp') to help planners quantify the welfare impact of investments in prediction versus other policy levers. Demonstrates its application in real-world case studies (German employment services, poverty targeting in Ethiopia) to derive context-specific conclusions.
73. TSJNet: A Multi-modality Target and Semantic Awareness Joint-driven Image Fusion Network
Core Problem: Unimodal images provide incomplete information for semantic segmentation and object detection tasks, and existing multimodal fusion methods have limited capability in discriminative modeling of multi-scale semantic structures and salient target regions.
Key Innovation: TSJNet, a novel multi-modality image fusion network that leverages semantic information from high-level tasks to guide the fusion process, using a multi-dimensional feature extraction module and a data-agnostic spatial attention module, along with a new multimodal UAV dataset (UMS).
74. Kernel-based Optimally Weighted Conformal Time-Series Prediction
Core Problem: Developing a robust conformal prediction method for time-series data that provides reliable conditional coverage guarantees for non-exchangeable data while achieving narrow prediction intervals.
Key Innovation: KOWCPI (Kernel-based Optimally Weighted Conformal Prediction Intervals), a novel conformal prediction method for time-series that adapts the Reweighted Nadaraya-Watson estimator and learns optimal data-adaptive weights, demonstrating superior performance in terms of narrower confidence intervals without losing coverage.
75. A High Resolution Urban and Rural Settlement Map of Africa Using Deep Learning and Satellite Imagery
Core Problem: Existing urban-rural settlement datasets for Africa are often spatially coarse, methodologically inconsistent, and poorly adapted to heterogeneous regions, limiting their usefulness for policy and research, especially for fine-scale analysis.
Key Innovation: A DeepLabV3-based deep learning framework integrating multi-source satellite imagery to produce a 10m resolution, accurate, and consistent urban-rural settlement map of Africa (HUR dataset), outperforming existing global products and providing an open framework for high-resolution settlement mapping.
76. ERVD: An Efficient and Robust ViT-Based Distillation Framework for Remote Sensing Image Retrieval
Core Problem: Efficient and robust retrieval of remote sensing images is crucial for various applications, but existing methods, especially those leveraging Vision Transformers (ViT), may lack optimal efficiency or robustness.
Key Innovation: Proposes ERVD, an Efficient and Robust ViT-Based Distillation Framework, designed to improve the performance and efficiency of remote sensing image retrieval, thereby enhancing the accessibility and utility of remote sensing data for analysis.
77. Latent Domain Modeling Improves Robustness to Geographic Shifts
Core Problem: Geographic distribution shift in geo-tagged datasets leads to uneven generalization across different spatially-determined groups, and existing domain adaptation methods often ignore available geographic coordinates.
Key Innovation: A general modeling framework that improves robustness to geographic distribution shift by modeling continuous, latent domain assignment using location encoders and conditioning the main task predictor on these jointly-trained latents, achieving significant improvements in worst-group performance on diverse geo-tagged image datasets.
78. Targetless LiDAR-Camera Calibration with Neural Gaussian Splatting
Core Problem: Traditional LiDAR-camera calibration methods rely on impractical physical targets and are prone to degradation over time, necessitating frequent recalibration for multi-sensor systems.
Key Innovation: Presents TLC-Calib, a targetless LiDAR-camera calibration method that jointly optimizes sensor poses with a neural Gaussian-based scene representation, achieving robust, generalizable, and consistent calibration without physical targets.
79. SIMSHIFT: A Benchmark for Adapting Neural Surrogates to Distribution Shifts
Core Problem: Neural surrogates for Partial Differential Equations (PDEs) suffer significant performance degradation when evaluated on problem configurations outside their training distribution, and Unsupervised Domain Adaptation (UDA) techniques are largely unexplored in complex engineering simulations.
Key Innovation: Introduction of SIMSHIFT, a novel benchmark dataset and evaluation suite for four industrial simulation tasks, and systematic evaluation of extended UDA methods on neural surrogates, demonstrating potential for robust modeling under distribution shifts in industrially relevant scenarios.
80. Interpretability and Generalization Bounds for Learning Spatial Physics
Core Problem: Despite promising applications of ML to scientific problems, there is a lack of rigorous quantification of accuracy, convergence rates, and generalization bounds, particularly regarding the critical role of the data's function space.
Key Innovation: Rigorous quantification of accuracy, convergence rates, and generalization bounds for ML models applied to linear differential equations, identifying the critical role of data function space, introducing a mechanistic interpretability lens via Green's function representations, and proposing a new cross-validation technique for physical systems.
81. Modulate and Reconstruct: Learning Hyperspectral Imaging from Misaligned Smartphone Views
Core Problem: Hyperspectral reconstruction (HSR) from RGB images is an ill-posed problem with limited accuracy due to severe spectral information loss, and existing approaches typically rely on a single RGB image.
Key Innovation: A novel multi-image-to-hyperspectral reconstruction (MI-HSR) framework leveraging a triple-camera smartphone system with spectral filters, enabling richer spectral observations and achieving significant accuracy improvements, supported by the new Doomer dataset.
82. CoBEVMoE: Heterogeneity-aware Feature Fusion with Dynamic Mixture-of-Experts for Collaborative Perception
Core Problem: Collaborative perception systems struggle with fusing heterogeneous observations from multiple agents due to differences in viewpoints and spatial positions, often overlooking perceptual diversity.
Key Innovation: CoBEVMoE, a novel collaborative perception framework that uses a Dynamic Mixture-of-Experts (DMoE) architecture in Bird's Eye View (BEV) space to explicitly model both feature similarity and heterogeneity across agents, enhanced by a Dynamic Expert Metric Loss (DEML), achieving state-of-the-art performance in BEV segmentation and 3D object detection.
83. Through the Perspective of LiDAR: A Feature-Enriched and Uncertainty-Aware Annotation Pipeline for Terrestrial Point Cloud Segmentation
Core Problem: Accurate semantic segmentation of terrestrial laser scanning (TLS) point clouds is limited by the costly and labor-intensive nature of manual annotation.
Key Innovation: Proposes a semi-automated, uncertainty-aware pipeline that integrates spherical projection, feature enrichment, ensemble learning, and targeted annotation to reduce labeling effort while sustaining high accuracy for TLS point cloud segmentation, demonstrated with the Mangrove3D dataset.
84. Breaking Scale Anchoring: Frequency Representation Learning for Accurate High-Resolution Inference from Low-Resolution Training
Core Problem: In Zero-Shot Super-Resolution Spatiotemporal Forecasting, models trained on low-resolution data suffer from 'Scale Anchoring,' where errors are anchored at low resolution because they cannot process unseen high-frequency components during high-resolution inference.
Key Innovation: Proposing architecture-agnostic Frequency Representation Learning (FRL) with resolution-aligned frequency representations and spectral consistency training, which alleviates Scale Anchoring, allowing errors to decrease with increasing resolution and significantly outperforming baselines in spatiotemporal forecasting.
85. SuperiorGAT: Graph Attention Networks for Sparse LiDAR Point Cloud Reconstruction in Autonomous Systems
Core Problem: LiDAR-based perception in autonomous systems is limited by fixed vertical beam resolution and compromised by beam dropout, leading to sparse point clouds with missing elevation information.
Key Innovation: SuperiorGAT, a graph attention-based framework, reconstructs missing elevation information in sparse LiDAR point clouds by modeling scans as beam-aware graphs and incorporating gated residual fusion, achieving lower reconstruction error and improved geometric consistency.
86. RealPDEBench: A Benchmark for Complex Physical Systems with Real-World Data
Core Problem: Scientific Machine Learning (ML) models for complex physical systems are bottlenecked by the lack of expensive real-world data, leading to models primarily trained on simulated data and hindering research into sim-to-real transfer.
Key Innovation: RealPDEBench, the first benchmark for scientific ML that integrates real-world measurements with paired numerical simulations across five datasets, three tasks, and eight metrics, facilitating the development of methods to bridge the sim-to-real gap.
87. G2P: Gaussian-to-Point Attribute Alignment for Boundary-Aware 3D Semantic Segmentation
Core Problem: Sparse and irregular point distributions in 3D point clouds provide limited appearance evidence, making geometry-only features insufficient to distinguish objects with similar shapes but distinct appearances in semantic segmentation.
Key Innovation: G2P (Gaussian-to-Point), a method that transfers appearance-aware attributes from 3D Gaussian Splatting to point clouds, establishing point-wise correspondences to resolve geometric ambiguity and leverage Gaussian opacity/scale attributes for more discriminative and boundary-aware 3D semantic segmentation.
88. Calibrated Multi-Level Quantile Forecasting
Core Problem: Existing quantile forecasting methods often lack guaranteed calibration at multiple quantile levels simultaneously, especially in online settings and against distribution shifts, while also maintaining quantile ordering.
Key Innovation: Development of MultiQT, an online method that guarantees calibrated multi-level quantile forecasts, maintains quantile ordering (e.g., 0.5-level never larger than 0.6-level), and has a no-regret guarantee, significantly improving calibration for various forecasting problems.
89. Toward Outdoor Population Presence Monitoring With Mobile Network Data and Satellite Imagery
Core Problem: The need for dynamic population mapping that differentiates indoor and outdoor activity to enhance accuracy for applications like smart city planning, emergency response, and public health, moving beyond traditional census data.
Key Innovation: A procedure for outdoor population detection that combines passively collected mobile network data with GPS data for spatial accuracy and validates results with satellite imagery of detected pedestrians, demonstrating strong potential for near real-time monitoring of population presence on streets.
90. Fusing ERA5-Land and SMAP L4 for an improved global soil moisture product (1950–2025)
Core Problem: Current global soil moisture products suffer from inconsistencies, coverage gaps, and biases, limiting their utility for hydrological modeling, climate studies, and ecosystem management.
Key Innovation: Development of an adjusted ERA5-Land dataset (1950-2025) by fusing ERA5-Land and SMAP L4 using a mean-variance rescaling method, resulting in a global soil moisture product with enhanced spatiotemporal coverage, reduced bias, and improved accuracy for applications like drought monitoring and water resource management.
91. Uncertainty, temporal variability, and influencing factors of empirical streamflow sensitivities
Core Problem: There are open questions regarding the robustness, temporal variability, and influencing factors of empirically-derived streamflow sensitivities to precipitation and potential evaporation, which are widely used for catchment characterization and climate change impact assessments.
Key Innovation: This paper re-investigates theoretical and empirical approaches for streamflow sensitivities, confirming multiple regression's superiority but highlighting high uncertainty for potential evaporation. It demonstrates that sensitivities are not static, decreasing significantly over time with increasing aridity index, and are influenced by catchment storage processes and data uncertainty, urging caution in their use for climate change impact assessments.
92. Hybrid SDF-CFD-DEM analysis of suffusion behavior in coral sand incorporating irregular particle morphology and intraparticle voids
Core Problem: Insufficient understanding of how irregular particle morphology and intraparticle voids in coral sand influence suffusion behavior, leading to inaccurate predictions with simplified models.
Key Innovation: Development of a hybrid SDF-CFD-DEM numerical framework that accurately accounts for irregular particle morphology and intraparticle voids, revealing that intraparticle voids inhibit suffusion and simplified spherical models significantly overestimate fine particle erosion.
93. Global rainfall simulator studies: review, challenges and perspectives
Core Problem: The lack of standardization and limitations in rainfall simulator technology hinder their effectiveness in accurately reproducing natural rainfall and extrapolating small-scale results to larger scales for hydrological and soil erosion research.
Key Innovation: Provides a comprehensive review of rainfall simulator technology, identifies key limitations (raindrop characteristics, spatiotemporal differences, scale effects), and proposes future directions including AI integration, comparative analyses, database development, multiscale validation, and establishing a comparability framework for standardization.
94. Gaussian-Constrained LeJEPA Representations for Unsupervised Scene Discovery and Pose Consistency
Core Problem: Unsupervised 3D scene reconstruction from unstructured image collections remains challenging, especially with multiple unrelated scenes, visual ambiguity, outliers, and mixed content, as highlighted by the IMC2025.
Key Innovation: An empirical evaluation of Gaussian-constrained representations inspired by LeJEPA, which enforce isotropic Gaussian constraints on learned image embeddings, demonstrating improved scene separation and pose plausibility in 3D reconstruction, particularly in visually ambiguous settings.
95. Guidestar-Free Adaptive Optics with Asymmetric Apertures
Core Problem: Existing adaptive optics (AO) systems typically require a guidestar or wavefront sensor to correct optical aberrations, limiting their application in scenarios without such references or in the presence of unknown obscurants.
Key Innovation: Develops the first closed-loop guidestar-free adaptive optics (AO) system that uses asymmetric apertures, machine learning for point-spread-function (PSF) estimation and phase reconstruction from natural scenes, and a spatial light modulator for real-time optical correction, outperforming state-of-the-art methods with significantly fewer measurements and less computation.
96. COMBOOD: A Semiparametric Approach for Detecting Out-of-distribution Data for Image Classification
Core Problem: Identifying out-of-distribution (OOD) data at inference time is crucial for many machine learning applications, but existing methods often struggle to perform accurately across both near-OOD and far-OOD scenarios.
Key Innovation: Presents COMBOOD, a novel unsupervised semi-parametric framework for OOD detection in image recognition that combines signals from nearest-neighbor and Mahalanobis distance metrics to derive an accurate confidence score, outperforming state-of-the-art methods on benchmark datasets for both near-OOD and far-OOD scenarios.
97. MTS-CSNet: Multiscale Tensor Factorization for Deep Compressive Sensing on RGB Images
Core Problem: Deep learning based compressive sensing (CS) methods typically learn sampling operators using convolutional or block-wise fully connected layers, which limit receptive fields and scale poorly for high-dimensional data.
Key Innovation: MTSCSNet, a CS framework based on Multiscale Tensor Summation (MTS) factorization, which uses MTS as a learnable CS operator and for reconstruction, enabling large receptive fields, efficient cross-dimensional correlation modeling, and state-of-the-art reconstruction performance on RGB images with a compact architecture.
98. Bidirectional Reward-Guided Diffusion for Real-World Image Super-Resolution
Core Problem: Diffusion-based super-resolution models trained on synthetic data perform poorly on real-world low-resolution images due to distribution shifts.
Key Innovation: Bird-SR, a bidirectional reward-guided diffusion framework that optimizes super-resolution via reward feedback learning, leveraging both synthetic and real-world images, and employing a dynamic fidelity-perception weighting strategy to balance structural and perceptual learning.
99. Zero-Shot UAV Navigation in Forests via Relightable 3D Gaussian Splatting
Core Problem: UAV navigation in unstructured outdoor environments (like forests) using passive monocular vision is challenging due to the substantial visual domain gap between simulation and reality, especially with dynamic lighting conditions.
Key Innovation: A novel end-to-end reinforcement learning framework combined with 'Relightable 3D Gaussian Splatting' enables zero-shot UAV navigation in complex forest environments, achieving robust, collision-free movement at high speeds and significant resilience to drastic lighting variations without fine-tuning.
100. Systematic Performance Assessment of Deep Material Networks for Multiscale Material Modeling
Core Problem: Despite their growing adoption, systematic evaluations of Deep Material Networks (DMNs) performance across the full offline-online pipeline (prediction accuracy, computational efficiency, training robustness, and generalization from linear elastic to nonlinear inelastic regimes) remain limited.
Key Innovation: Presents a comprehensive comparative assessment of DMNs, investigating the effects of offline training choices on online generalization performance and uncertainty. It demonstrates that the rotation-free Interaction-based Material Network (IMN) formulation achieves significant speed-up in training while maintaining comparable accuracy and efficiency.
101. Graph homophily booster: Reimagining the role of discrete features in heterophilic graph learning
Core Problem: Existing Graph Neural Networks (GNNs) struggle with heterophilic graphs, where connected nodes have dissimilar features or labels, often performing worse than simple MLPs on challenging datasets, and current solutions primarily focus on architectural designs without directly addressing the root cause of heterophily.
Key Innovation: Proposes GRAPHITE, a framework that directly increases graph homophily by transforming the graph. It creates feature nodes to facilitate homophilic message passing between nodes with similar features, significantly boosting homophily and outperforming state-of-the-art methods on heterophilic graphs while maintaining accuracy on homophilic ones.
102. FEM-Informed Hypergraph Neural Networks for Efficient Elastoplasticity
Core Problem: Developing efficient and accurate physics-informed machine learning models for complex elastoplastic problems in computational mechanics, overcoming limitations of existing PINN variants.
Key Innovation: Proposes FEM-Informed Hypergraph Neural Networks (FHGNN) that embed FEM computations directly into message-passing layers, using a physics-driven, label-free training approach with an efficient variational loss, achieving improved accuracy and efficiency for large 3D elastoplastic problems.
103. Active Learning Using Aggregated Acquisition Functions: Accuracy and Sustainability Analysis
Core Problem: Active learning faces the exploration-exploitation dilemma, where representativity-based functions explore but do not prioritize boundary decisions, and uncertainty-based functions refine boundaries but do not explore, leading to pathologies like batch mode inefficiency and the cold start problem.
Key Innovation: Introduces six aggregation structures (series, parallel, hybrid, adaptive feedback, random exploration, annealing exploration) for acquisition functions to alleviate AL pathologies, balancing accuracy and energy consumption, demonstrating robust results and reduced computational costs.
104. Thermal odometry and dense mapping using learned ddometry and Gaussian splatting
Core Problem: Existing thermal odometry and mapping approaches are predominantly geometric, often fail across diverse datasets, and lack the ability to produce dense maps, limiting their utility for robust environmental perception in adverse conditions.
Key Innovation: Proposes TOM-GS, a thermal odometry and mapping method that integrates learning-based odometry with Gaussian Splatting (GS)-based dense mapping, featuring dedicated thermal image enhancement and monocular depth integration, achieving superior performance in motion estimation and dense reconstruction in adverse conditions.
105. Revealing the Semantic Selection Gap in DINOv3 through Training-Free Few-Shot Segmentation
Core Problem: While self-supervised Vision Transformers like DINOv3 provide rich features for dense vision tasks, there's a 'Semantic Selection Gap' where traditional heuristics fail to reliably identify the globally optimal intermediate feature representations for tasks like few-shot semantic segmentation, despite the last-layer being a strong baseline.
Key Innovation: Establishes FSSDINO, a training-free baseline for few-shot semantic segmentation using frozen DINOv3 features, and through Oracle-guided layer analysis, reveals a significant performance gap between standard last-layer features and globally optimal intermediate representations, characterizing the 'Semantic Selection Gap'.
106. Enhancing Time Series Classification with Diversity-Driven Neural Network Ensembles
Core Problem: Most existing neural network-based ensemble methods for Time Series Classification (TSC) train models with identical configurations, leading to redundant feature representations and limiting the benefits of ensembling.
Key Innovation: A diversity-driven ensemble learning framework that explicitly encourages feature diversity among neural network ensemble members using a feature orthogonality loss, achieving state-of-the-art performance with fewer models and improved efficiency and scalability.
107. Escaping Spectral Bias without Backpropagation: Fast Implicit Neural Representations with Extreme Learning Machines
Core Problem: Training Implicit Neural Representations (INRs) to capture fine-scale details is hindered by iterative backpropagation, spectral bias, and difficulty with highly non-uniform frequency content.
Key Innovation: ELM-INR, a backpropagation-free INR that decomposes the domain into subdomains and uses Extreme Learning Machines (ELM) for closed-form local fitting. It also introduces BEAM, an adaptive mesh refinement strategy, to improve reconstruction quality by balancing spectral complexity.
108. Quantifying Explanation Quality in Graph Neural Networks using Out-of-Distribution Generalization
Core Problem: Current evaluation metrics for post-hoc explanations in Graph Neural Networks (GNNs) often fail to assess whether an explanation identifies the true underlying causal variables, hindering the understanding and trustworthiness of GNN models.
Key Innovation: Proposes the Explanation-Generalization Score (EGS), a metric that quantifies the causal relevance of GNN explanations by evaluating their stability across distribution shifts (Out-of-Distribution generalization), providing a principled benchmark for ranking explainers based on their ability to capture causal substructures.
109. Exploiting Free-Surface Ghosts as Mirror Observations in Marine Seismic Data
Core Problem: Free-surface ghosts in marine seismic data are traditionally treated as artifacts that degrade bandwidth and temporal resolution, requiring mitigation through acquisition design or inverse filtering, which can be numerically unstable.
Key Innovation: Proposes a processing-driven framework that reinterprets free-surface ghosts as coherent mirror observations, physically realigning and summing primary and ghost wavefields to enhance signal quality, improve wavelet compactness, and partially recover ghost-affected frequency content without explicit inversion.
110. DINO-Mix: Distilling Foundational Knowledge with Cross-Domain CutMix for Semi-supervised Class-imbalanced Medical Image Segmentation
Core Problem: Prevailing semi-supervised learning (SSL) frameworks for medical image segmentation suffer from confirmation bias under class imbalance, leading to poor recognition of minority classes.
Key Innovation: DINO-Mix is a multi-level 'outward-looking' framework that uses Foundational Knowledge Distillation from a pre-trained visual foundation model (DINOv3) as an unbiased external semantic teacher, combined with Progressive Imbalance-aware CutMix (PIC), to address class imbalance in semi-supervised image segmentation.
111. Nansde-net: A neural sde framework for generating time series with memory
Core Problem: Modeling time series with long- or short-memory characteristics is challenging, and existing approaches like fractional Brownian motion are incompatible with It
Key Innovation: Introduces NANSDE-Net, a generative model that extends Neural SDEs by incorporating Neural Network-kernel ARMA-type noise (NA-noise), an It
112. Interpretable Dynamic Network Modeling of Tensor Time Series via Kronecker Time-Varying Graphical Lasso
Core Problem: Modeling dynamic network structures in real-world tensor time series is challenging due to large, entangled networks that are hard to interpret and computationally intensive to estimate.
Key Innovation: Proposes Kronecker Time-Varying Graphical Lasso (KTVGL), a method that estimates mode-specific dynamic networks in a Kronecker product form, leading to interpretable results and computational efficiency, even for stream algorithms, demonstrated on synthetic and real-world data.
113. Distribution-Free Robust Functional Predict-Then-Optimize
Core Problem: Neural operator surrogate models, while computationally efficient for solving PDEs in decision-making, fail to provide calibrated notions of uncertainty in their predictions, and existing approaches either require restrictive distributional assumptions or lack scalability.
Key Innovation: A novel application of conformal prediction to produce distribution-free uncertainty quantification over the function spaces mapped by neural operators, enabling formal regret characterization in downstream robust decision-making tasks, and demonstrating efficient solvability using an infinite-dimensional generalization of Danskin's Theorem.
114. Tighnari v2: Mitigating Label Noise and Distribution Shift in Multimodal Plant Distribution Prediction via Mixture of Experts and Weakly Supervised Learning
Core Problem: Large-scale plant distribution prediction faces challenges due to sparse/biased observational data (PA/PO), label noise in PO data, and significant geographic distribution shifts between training/test samples, hindering accurate biodiversity conservation efforts.
Key Innovation: Tighnari v2, a multimodal fusion framework that leverages both PA and PO data with a pseudo-label aggregation strategy, a novel model architecture (Swin Transformer, TabM, Temporal Swin Transformer with tri-modal cross-attention), and a mixture-of-experts paradigm to mitigate label noise and distribution shifts.
115. Radial M\"untz-Sz\'asz Networks: Neural Architectures with Learnable Power Bases for Multidimensional Singularities
Core Problem: Coordinate-separable neural architectures struggle to accurately model radial singular fields (e.g., 1/r, log r, crack-tip profiles) due to fundamental mathematical obstructions.
Key Innovation: Introduces Radial M"untz-Sz"asz Networks (RMN), neural architectures that represent fields as linear combinations of learnable radial powers r^μ and a limit-stable log-primitive, enabling accurate modeling of multidimensional singularities, closed-form spatial gradients, and physics-informed learning on punctured domains with significantly lower RMSE.
116. Learning Credal Ensembles via Distributionally Robust Optimization
Core Problem: Existing credal predictors primarily quantify epistemic uncertainty (EU) based on sensitivity to optimization randomness, failing to capture deeper sources of uncertainty arising from potential distribution shifts between training and test data.
Key Innovation: Defines EU as disagreement among models trained with varying relaxations of the i.i.d. assumption. Proposes CreDRO, a framework that learns an ensemble of plausible models via distributionally robust optimization, capturing EU from both training randomness and meaningful disagreement due to distribution shifts, improving robustness in OOD detection and selective classification.
117. OneVision-Encoder: Codec-Aligned Sparsity as a Foundational Principle for Multimodal Intelligence
Core Problem: Modern vision architectures process dense pixel grids uniformly, wasting compute on redundant information instead of focusing on sparse, discriminative information (motion, meaning), hindering efficient and accurate multimodal intelligence.
Key Innovation: OneVision-Encoder aligns vision architectures with codec principles by using Codec Patchification to focus computation on signal-rich regions (3.1%-25% of pixels), unifying spatial and temporal reasoning, leading to improved efficiency and accuracy in multimodal understanding.
118. Foundation Inference Models for Ordinary Differential Equations
Core Problem: Inferring vector fields of Ordinary Differential Equations (ODEs) from noisy trajectory data is challenging, often requiring complex training pipelines, machine learning expertise, or strong system-specific prior knowledge.
Key Innovation: FIM-ODE, a pretrained Foundation Inference Model that amortizes low-dimensional ODE inference by directly predicting the vector field from noisy trajectory data in a single forward pass, offering strong zero-shot performance and fast, stable adaptation for finetuning.
119. Redundancy-Free View Alignment for Multimodal Human Activity Recognition with Arbitrarily Missing Views
Core Problem: Multimodal multiview learning approaches struggle with flexible view configurations, such as arbitrary view combinations, varying numbers of views, and heterogeneous modalities, especially when views are missing.
Key Innovation: RALIS, a model combining multiview contrastive learning with a mixture-of-experts module, which uses an adjusted center contrastive loss for redundancy-free view alignment and self-supervised representation learning, and a load-balancing mixture-of-experts module to adapt to arbitrary view combinations, supporting arbitrarily missing views during training and inference.
120. FreqLens: Interpretable Frequency Attribution for Time Series Forecasting
Core Problem: Time series forecasting models often lack interpretability, limiting their adoption in domains requiring explainable predictions.
Key Innovation: Introduces FreqLens, an interpretable forecasting framework with learnable frequency discovery (parameterized frequency bases learned from data) and axiomatic frequency attribution (theoretically grounded framework providing per-frequency Shapley values), achieving competitive performance while discovering physically meaningful frequencies.
121. TiFRe: Text-guided Video Frame Reduction for Efficient Video Multi-modal Large Language Models
Core Problem: Video Multi-Modal Large Language Models (Video MLLMs) face high computational costs from processing numerous video frames, and simple fixed-rate frame selection often overlooks valuable information, leading to performance degradation.
Key Innovation: Text-guided Video Frame Reduction (TiFRe), a framework that uses a Text-guided Frame Sampling (TFS) strategy to select key frames based on user input and a Frame Matching and Merging (FMM) mechanism to integrate non-key frame information into selected key frames, reducing computational costs while preserving video semantics.
122. Distributionally Robust Optimization via Generative Ambiguity Modeling
Core Problem: Developing an effective ambiguity set for Distributionally Robust Optimization (DRO) that is consistent with the nominal distribution, diverse enough for various scenarios, and leads to tractable solutions.
Key Innovation: Proposes generative model-based ambiguity sets that capture adversarial distributions beyond the nominal support space while maintaining consistency. It introduces GAS-DRO (DRO with Generative Ambiguity Set), a tractable algorithm that solves the inner maximization over the parameterized generative model space, demonstrating superior Out-of-Distribution (OOD) generalization performance.
123. Improving Detection of Rare Nodes in Hierarchical Multi-Label Learning
Core Problem: In hierarchical multi-label classification, it is challenging to enable model predictions to reach deeper levels of the hierarchy for more detailed classifications, primarily due to the natural rarity of certain classes (nodes) and hierarchical constraints.
Key Innovation: Proposes a weighted loss objective for neural networks that combines node-wise imbalance weighting with focal weighting components, leveraging modern quantification of ensemble uncertainties. This approach emphasizes rare and uncertain nodes, leading to significant improvements in recall (up to five-fold) and F1 score on benchmark datasets, even with suboptimal encoders or limited data.
124. ShapeCond: Fast Shapelet-Guided Dataset Condensation for Time Series Classification
Core Problem: Existing dataset condensation methods are image-centric and fail to capture time-series-specific temporal structures like shapelets, leading to inefficiency and suboptimal accuracy for time series classification.
Key Innovation: Proposes ShapeCond, a novel and efficient time series dataset condensation framework that uses a shapelet-guided optimization strategy to synthesize compact training sets, improving accuracy and speed.
125. NLP Sampling: Combining MCMC and NLP Methods for Diverse Constrained Sampling
Core Problem: Generating diverse samples under hard constraints is a significant challenge across many domains, and there's a need for an integrative framework combining different computational methods.
Key Innovation: Proposes "NLP Sampling" as a general problem formulation and a family of restarting two-phase methods to integrate MCMC, constrained optimization, and robotics techniques for diverse constrained sampling, providing insights into their strengths.
126. Graph-Based Nearest-Neighbor Search without the Spread
Core Problem: Existing nearest-neighbor graph constructions for approximate nearest-neighbor (ANN) queries depend on the "spread" parameter, which can be unbounded, limiting their practical applicability.
Key Innovation: Develops an external linear-size data structure that, when combined with a linear-size nearest-neighbor graph, allows for answering ANN queries in logarithmic time in the number of points, effectively removing the dependency on the spread.
127. Curriculum-Learned Vanishing Stacked Residual PINNs for Hyperbolic PDE State Reconstruction
Core Problem: Modeling distributed dynamical systems governed by hyperbolic partial differential equations (PDEs) is challenging due to discontinuities and shocks, which impede the convergence of traditional physics-informed neural networks (PINNs).
Key Innovation: Integrates three curriculum-learning methods (primal-dual optimization, causality progression, and adaptive sampling) into the vanishing stacked residual PINN (VSR-PINN). This approach systematically reduces point-wise MSE and its variability, significantly improving the reconstruction of hyperbolic PDE states by enabling a smooth transition from parabolic to hyperbolic regimes and addressing discontinuities.
128. AI for Sustainable Data Protection and Fair Algorithmic Management in Environmental Regulation
Core Problem: Ensuring sustainable data protection and fair algorithmic management for environmental data in the era of evolving cyber threats, given the limitations of traditional encryption methods.
Key Innovation: Evaluating how AI can enhance homomorphic encryption and multi-party computation for robust environmental data protection, highlighting AI-driven dynamic key management, adaptive encryption schemes, and optimized computational efficiency.
129. Informative Object-centric Next Best View for Object-aware 3D Gaussian Splatting in Cluttered Scenes
Core Problem: In cluttered scenes, existing 3D Gaussian Splatting (3DGS) methods for selecting Next Best Views (NBV) rely solely on geometric cues, neglect manipulation-relevant semantics, and do not prioritize underexplored regions, leading to incomplete or unreliable 3D representations.
Key Innovation: An instance-aware and object-centric Next Best View (NBV) policy for 3D Gaussian Splatting that leverages object features and confidence-weighted information gain to prioritize underexplored regions and specific target objects, significantly reducing depth error and improving reconstruction robustness in cluttered scenes.
130. LBL: Logarithmic Barrier Loss Function for One-class Classification
Core Problem: Despite advances in one-class classification (OCC), there is a lack of effective deep learning loss functions that can derive more compact hyperspheres and ensure stable optimization.
Key Innovation: Introduction of LBL (Logarithmic Barrier Loss) and LBLSig (LBL with unilateral relaxation Sigmoid function) as novel OCC loss functions. LBL assigns large gradients to margin samples for compact hyperspheres, and LBLSig addresses instability by fusing MSE and CE, leading to smoother optimization.
131. Towards Transparent and Efficient Anomaly Detection in Industrial Processes through ExIFFI
Core Problem: Conventional anomaly detection methods in industrial settings merely label observations as normal or anomalous, lacking crucial insights and interpretable outcomes desirable for understanding model decisions in Industry 5.0.
Key Innovation: ExIFFI, the first industrial application of an approach for fast, efficient explanations for the Extended Isolation Forest (EIF) anomaly detection method, demonstrating superior explanation effectiveness, computational efficiency, and improved raw AD performance.
132. Disentangled Parameter-Efficient Linear Model for Long-Term Time Series Forecasting
Core Problem: Long-Term Time Series Forecasting (LTSF) suffers from complex deep models overfitting and linear models having quadratic parameter redundancy and entangled temporal/frequential properties, leading to inefficiency and suboptimal performance.
Key Innovation: DiPE-Linear, a novel disentangled parameter-efficient linear model for LTSF that separates the monolithic weight matrix into specialized modules (Static Frequential Attention, Static Time Attention, Independent Frequential Mapping) with low-rank weight sharing, reducing complexity and achieving state-of-the-art performance.
133. Probabilistic Forecasting via Autoregressive Flow Matching
Core Problem: Probabilistic forecasting of multivariate timeseries data requires models that can capture complex multi-modal conditional distributions while retaining benefits like strong extrapolation performance, compact model size, and well-calibrated uncertainty estimates.
Key Innovation: FlowTime, a generative model that formulates forecasting as sampling from a learned conditional distribution over future trajectories, decomposing the joint distribution into a sequence of conditional densities modeled via a shared flow using the flow matching (FM) framework, enabling scalable and simulation-free learning.
134. VisionReasoner: Unified Reasoning-Integrated Visual Perception via Reinforcement Learning
Core Problem: Large vision-language models need a unified framework capable of reasoning and solving diverse visual perception tasks (detection, segmentation, counting) within a shared model efficiently.
Key Innovation: Introduces VisionReasoner, a unified framework that enhances reasoning capabilities through a unified reward mechanism and multi-object cognitive learning strategies, enabling it to address diverse perception tasks and generate faithful reasoning processes, outperforming baselines.
135. PiFlow: Principle-Aware Scientific Discovery with Multi-Agent Collaboration
Core Problem: Existing LLM-based multi-agent systems for scientific discovery often lack rationality constraints and struggle to consistently link hypotheses with evidence, leading to inefficient and aimless exploration.
Key Innovation: Introduces PiFlow, an information-theoretical framework that treats automated scientific discovery as a structured uncertainty reduction problem guided by scientific principles. It significantly improves discovery efficiency, solution quality, and speed-to-solution across scientific domains.
136. Toward Inherently Robust VLMs Against Visual Perception Attacks
Core Problem: Deep neural networks in autonomous vehicles are vulnerable to visual perception attacks, leading to misclassification and safety threats, with existing defenses often failing to generalize and degrading clean accuracy.
Key Innovation: Vehicle Vision-Language Models (V2LMs), fine-tuned for autonomous vehicle perception, demonstrate inherent robustness to unseen attacks without adversarial training, maintaining substantially higher adversarial accuracy (declining by under 8% on average) compared to conventional DNNs.
137. RealSR-R1: Reinforcement Learning for Real-World Image Super-Resolution with Vision-Language Chain-of-Thought
Core Problem: Existing Real-World Image Super-Resolution methods struggle with accurately understanding degraded image content, leading to reconstructed results that are low-fidelity and unnatural.
Key Innovation: RealSR-R1, which empowers RealSR models with understanding and reasoning capabilities through a Vision-Language Chain-of-Thought (VLCoT) framework and introduces Group Relative Policy Optimization (GRPO) with four reward functions to generate realistic details and accurately understand image content, particularly in semantically rich or severely degraded scenes.
138. mTSBench: Benchmarking Multivariate Time Series Anomaly Detection and Model Selection at Scale
Core Problem: Anomaly detection in multivariate time series is challenging due to high-dimensional dependencies, cross-correlations, and scarce labeled anomalies, compounded by a lack of comprehensive benchmarks and effective model selection strategies.
Key Innovation: Introduction of mTSBench, the largest benchmark for multivariate time series anomaly detection and model selection, providing a comprehensive evaluation of 24 detectors and three model selection methods, highlighting the need for robust selection strategies.
139. Robust Image Stitching with Optimal Plane
Core Problem: Existing image stitching methods often lack robustness across diverse real-world scenes and struggle to balance content alignment with structural preservation, leading to unnatural results.
Key Innovation: RopStitch, an unsupervised deep image stitching framework that incorporates a dual-branch architecture (pretrained for semantic invariance, learnable for fine-grained features) for robustness, and a concept of virtual optimal planes (modeled as homography decomposition coefficients with an iterative predictor and minimal semantic distortion constraint) to resolve the conflict between content alignment and structural preservation.
140. Hyperspectral Imaging
Core Problem: The need for a comprehensive overview of Hyperspectral Imaging (HSI) technology, its underlying principles, data acquisition, analysis methods, diverse applications, and persistent challenges.
Key Innovation: A comprehensive primer on HSI, summarizing its physical principles, sensor architectures, data processing, classical and AI-driven analysis methods, diverse applications (including Earth observation), challenges, and emerging solutions, highlighting its potential as a general-purpose, cross-disciplinary platform for advanced monitoring and decision-making.
141. Residual Vector Quantization For Communication-Efficient Multi-Agent Perception
Core Problem: Multi-agent collaborative perception improves scene understanding but is constrained by communication bandwidth, limiting scalability and efficient information sharing across connected agents.
Key Innovation: Presents ReVQom, a learned feature codec that preserves spatial identity while compressing intermediate features via multi-stage residual vector quantization (RVQ), reducing payloads from 8192 bits per pixel (bpp) to 6-30 bpp with minimal accuracy loss, enabling efficient and accurate multi-agent collaborative perception.
142. Vid-LLM: A Compact Video-based 3D Multimodal LLM with Reconstruction-Reasoning Synergy
Core Problem: Extending Multimodal Large Language Models (MLLMs) to 3D scene understanding remains a major challenge, as existing 3D-MLLMs often depend on 3D data inputs, limiting scalability and generalization.
Key Innovation: Proposes Vid-LLM, a video-based 3D-MLLM that directly processes video inputs without requiring external 3D data. It integrates geometric priors compactly using a Cross-Task Adapter (CTA) module, introduces a Metric Depth Model for real-scale geometry recovery, and uses a two-stage distillation optimization strategy, demonstrating superior multi-task capabilities in 3D Question Answering, 3D Dense Captioning, and 3D Visual Grounding tasks.
143. Probabilistic bias adjustment of seasonal predictions of Arctic Sea Ice Concentration
Core Problem: Seasonal prediction systems for Arctic sea ice concentration often exhibit systematic biases and complex spatio-temporal errors, and existing deterministic bias correction methods are limited in quantifying uncertainty and generating large ensembles.
Key Innovation: Introduces a probabilistic error correction framework based on a conditional Variational Autoencoder (VAE) model to map the conditional distribution of observations given biased model predictions, enabling the generation of large ensembles of adjusted forecasts with better calibration and smaller errors.
144. Generalised Flow Maps for Few-Step Generative Modelling on Riemannian Manifolds
Core Problem: Current geometric generative models on Riemannian manifolds are computationally expensive at inference, requiring many steps of complex numerical simulation.
Key Innovation: Generalised Flow Maps (GFM), a new class of few-step generative models that extends the Flow Map framework to arbitrary Riemannian manifolds. It unifies and elevates existing Euclidean few-step generative models and achieves state-of-the-art sample quality and competitive log-likelihoods on geometric datasets, including geospatial data.
145. Open-Set Domain Adaptation Under Background Distribution Shift: Challenges and A Provably Efficient Solution
Core Problem: Maintaining machine learning model performance in real-world scenarios where data shifts occur, specifically when new classes emerge (open-set recognition) and the distribution of known categories changes (background distribution shift).
Key Innovation: Developing CoLOR, a provably efficient method that is guaranteed to solve open-set recognition even under background distribution shift, outperforming existing methods on image and text data by making it scalable and robust.
146. ALIGN: Advanced Query Initialization with LiDAR-Image Guidance for Occlusion-Robust 3D Object Detection
Core Problem: Existing query initialization strategies in camera and LiDAR-based 3D object detection methods lead to inefficient query usage and reduced accuracy, particularly for occluded or crowded objects.
Key Innovation: Proposes ALIGN, an occlusion-robust, object-aware query initialization approach with three components: Occlusion-aware Center Estimation (OCE) integrating LiDAR geometry and image semantics, Adaptive Neighbor Sampling (ANS) for object candidates, and Dynamic Query Balancing (DQB) for query distribution.
147. TS-Arena -- A Live Forecast Pre-Registration Platform
Core Problem: Reliably evaluating Time Series Foundation Models (TSFMs) is difficult due to risks of train-test sample overlaps and temporal leakage when using historical data.
Key Innovation: TS-Arena, a live forecasting platform with a strict pre-registration protocol, requires models to submit predictions before ground-truth data exists, preventing test-set contamination and enabling robust longitudinal evaluation of forecasting models.
148. ThermoSplat: Cross-Modal 3D Gaussian Splatting with Feature Modulation and Geometry Decoupling
Core Problem: Extending 3D Gaussian Splatting (3DGS) to multi-spectral scenarios (RGB and thermal) is challenging due to difficulties in leveraging complementary information, neglecting cross-modal correlations, or failing to adaptively handle structural correlations and physical discrepancies between spectrums.
Key Innovation: Proposes ThermoSplat, a framework for deep spectral-aware reconstruction using active feature modulation (Spectrum-Aware Adaptive Modulation) and adaptive geometry decoupling (Modality-Adaptive Geometric Decoupling) to integrate RGB and thermal data, achieving state-of-the-art rendering quality.
149. Zero-shot Generalizable Graph Anomaly Detection with Mixture of Riemannian Experts
Core Problem: Existing zero-shot generalist Graph Anomaly Detection (GAD) methods largely ignore intrinsic geometric differences across diverse anomaly patterns, limiting their cross-domain generalization because a single static curvature space cannot capture geometry-dependent graph anomaly patterns.
Key Innovation: Proposes GAD-MoRE, a novel framework for zero-shot Generalizable Graph Anomaly Detection with a Mixture of Riemannian Experts architecture. It employs specialized Riemannian expert networks in distinct curvature spaces, an anomaly-aware multi-curvature feature alignment module, and a memory-based dynamic router to adaptively assign inputs to the most compatible expert, significantly outperforming baselines in zero-shot settings.
150. Predictability Enables Parallelization of Nonlinear State Space Models
Core Problem: The challenge of efficiently parallelizing nonlinear state space models, where the factors governing the difficulty of the corresponding optimization problems for parallel evaluation remained unclear.
Key Innovation: Establishing a precise relationship between a system's predictability (quantified by the largest Lyapunov exponent) and the conditioning of its parallelization optimization problem, showing that predictable systems can be computed in O((log T)^2) time, providing practical guidance for efficient parallelization.
151. Provable FDR Control for Deep Feature Selection: Deep MLPs and Beyond
Core Problem: There is a lack of theoretical guarantees for False Discovery Rate (FDR) control in feature selection within a general deep learning setting, limiting the reliability of feature importance assessments.
Key Innovation: Development of a flexible deep neural network-based feature selection framework that approximately controls the False Discovery Rate (FDR), applicable to various deep learning architectures, with theoretical guarantees for asymptotic FDR control.
152. A deep learning method for spatiotemporal significant wave height estimation with ship attitude compensation
Core Problem: Traditional significant wave height (SWH) sensors are costly and inflexible, while shipborne vision-based techniques suffer from accuracy degradation due to ship motion and dynamic disturbances.
Key Innovation: Presented an attitude-aware spatiotemporal deep learning framework that fuses sequential ocean surface images with synchronized ship attitude data and employs a multi-head self-attention mechanism to jointly capture wave dynamics and compensate for vessel-induced motion, achieving high predictive accuracy and robustness for real-time SWH estimation.
153. MSFE-Mamba: Multiscale Frequency-Enhanced Mamba for Hyperspectral Image Classification
Core Problem: Existing hyperspectral image classification approaches are limited by convolutional neural networks' short-range spatial context, transformers' quadratic complexity on high-dimensional spectra, and underutilization of the frequency domain, restricting effectiveness in complex scenes.
Key Innovation: A multiscale frequency-enhanced Mamba (MSFE-Mamba) unified framework that jointly models spatial, spectral, and frequency-domain information through novel modules for adaptive multiscale fusion, spectrofrequency modulation, and frequency-enhanced Mamba, achieving state-of-the-art classification accuracy.
154. MSSTFormer: Multiscale Spectral–Spatial Supertoken Aggregation Transformer for Hyperspectral Image Classification
Core Problem: Existing Transformer-based hyperspectral image (HSI) classification methods lack mechanisms to reinforce token representations, leading to redundant features and limited discrimination, and struggle to capture both local spatial information and global spectral changes across different scales.
Key Innovation: A novel multiscale spectral–spatial supertoken aggregation transformer (MSSTFormer) that introduces 'Supertoken' for compact feature aggregation, enhances spectral–spatial correlation modeling using a chained dual-attention mechanism, and integrates multiscale features for pixelwise classification, achieving higher accuracy.
155. A Semantic Segmentation of Remote Sensing Images via Spatial–Frequency Encoding and Multiscale Feature Fusion
Core Problem: Current mainstream methods for semantic segmentation of remote sensing images (CNNs, Transformers) face a tradeoff between effectively capturing long-range dependencies and maintaining computational efficiency, limiting performance in applications like land cover mapping.
Key Innovation: Introduction of WE-MambaNet, an innovative semantic segmentation model for RS images that combines spatial-frequency collaborative encoding with multiscale feature fusion, utilizing a dual-encoder architecture with wavelet-guided frequency analysis and an efficient Mamba-based decoding pathway, achieving superior performance on RS datasets.
156. Evaluating the Performance of China’s Tianmu-1 GNSS-R Satellite Constellation Observations in Land and Ocean Remote Sensing Applications
Core Problem: The need to evaluate the performance and utility of the newly launched Tianmu-1 GNSS-R satellite constellation for various land and ocean remote sensing applications.
Key Innovation: First comprehensive evaluation of 10 months of TM-1 GNSS-R observations, demonstrating excellent performance in retrieving soil moisture, freeze-thaw, waterbody, sea ice, sea surface winds, and significant wave height, establishing it as a valuable tool for Earth observation.
157. Adaptive Hybrid-Domain Feature Compensation and Selective Modeling for Feature Refinement Network for Fine-Grained Remote Sensing Segmentation
Core Problem: Semantic segmentation of remote sensing imagery is challenged by scene complexity, multiscale targets, intraclass heterogeneity, and interclass homogeneity, with conventional methods often undervaluing frequency variance.
Key Innovation: Introduction of AHSMNet, a multifrequency-aware network integrating an adaptive hybrid-domain feature compensation (ADFC) module for fine-grained semantic details and a weighted cross-Mamba (WECM) module for selective feature refinement, demonstrably augmenting boundary delineation and feature extraction efficacy.
158. Dual-Domain Masked Representation Learning for Semantic Segmentation of Remote Sensing Images
Core Problem: Existing self-supervised learning methods for remote sensing semantic segmentation primarily focus on the spatial domain, neglecting critical frequency information prevalent in remote sensing images, which limits the robustness and discriminative power of learned features.
Key Innovation: Introduction of the Dual-Domain Masked Representation (DDMR) learning framework, which jointly leverages spatial and randomized frequency masking, along with an amplitude-phase loss, to enhance the robustness and discriminative power of learned features for semantic segmentation in complex remote sensing scenarios.
159. CosmDiff: Integrating Multitemporal Optical-SAR Data With Conditional Diffusion Models for Optical Satellite Time Series Reconstruction
Core Problem: Persistent cloud cover leads to significant missing data in optical satellite time series, diminishing data quality and hindering monitoring of vegetation dynamics and land surface changes.
Key Innovation: CosmDiff, a novel framework that reconstructs optical satellite time series by integrating multimodal, multitemporal optical and SAR data using conditional diffusion models, featuring a Transformer-based network with a dimensional decomposition attention mechanism and providing uncertainty estimates.
160. Visible-Light-Guided Infrared Image Super Resolution With Dual Amplitude-Phase Optimization
Core Problem: Infrared cameras suffer from inherent low spatial resolution and complex imaging degradation, limiting image quality for critical applications like search-and-rescue and fire monitoring, with existing methods often neglecting unique infrared modality characteristics.
Key Innovation: vap-SR, a novel framework for visible-light-guided infrared image super-resolution that leverages conditional diffusion and dual amplitude-phase optimization to effectively compensate for infrared image deficiencies by exploiting rich structural priors from visible images, improving both reconstruction quality and downstream object detection.
161. Multivariate Gaussian process regression for characterization of geo-data spatial variability from limited and non-co-located measurements
Core Problem: Accurately characterizing geo-data spatial variability from limited and non-co-located multivariate measurements, as existing MGPR methods require extensive co-located data.
Key Innovation: Proposed a novel MGPR method that fuses limited and non-co-located multivariate measurements by representing spatial correlation structures using orthonormal basis functions in a data-driven manner, allowing joint prediction of multiple geo-data profiles with quantified uncertainty.
162. EAV-DETR: Efficient Arbitrary-View oriented object detection with probabilistic guarantees for UAV imagery
Core Problem: Existing oriented object detectors struggle with unique challenges in UAV imagery (substantial scale variations, dense clustering, arbitrary orientations) and lack probabilistic guarantees required for safety-critical applications.
Key Innovation: EAV-DETR, an efficient oriented object detection transformer for UAV imagery, featuring a scale-adaptive center supervision (SACS) strategy for enhanced feature representations, an anisotropic decoupled rotational attention (ADRA) module for superior feature alignment, and a pose-aware Mondrian conformal prediction (PA-MCP) method for conditional coverage guarantees and reliable uncertainty quantification, demonstrating improved AP and faster inference speed on aerial imagery datasets.
163. Formation mechanism and evolution of crystalline deposits in drainage systems of high-temperature water gushing tunnels
Core Problem: Crystallization blockage in drainage systems of high-temperature water gushing tunnels poses a long-term safety hazard by increasing the risk of lining cracks due to accumulated hot water, high pressure, and thermal stress.
Key Innovation: The study reveals that high water temperatures significantly accelerate calcium carbonate crystallization and ion leaching from shotcrete, promoting deposit formation and mutual transformation between calcite and aragonite, providing essential insights for preventing blockage in tunnel drainage systems.
164. Late Triassic Hydroclimatic Changes in Central China Linked to Evolving Mountain Topography
Core Problem: Understanding the causal linkage between the Late Triassic arid-to-humid climatic shift in South China and the orogenic uplift of the Qinling Orogen.
Key Innovation: Constrains the arid-to-humid climatic shift to ca. 228-207 Ma, uses detrital zircon geochemistry to show crustal thickening and significant mountain building (up to 5,000m) in the Qinling Orogen, and employs climate modeling to demonstrate the significant orographic effect on regional precipitation.
165. Machine learning enhanced data assimilation framework for multiscale carbonate rock characterization
Core Problem: Efficiently characterizing multiscale carbonate rock structures and multiphase flow behaviors, overcoming the resource-intensive nature of multi-scale imaging and the computational cost of numerical simulations.
Key Innovation: Proposes a machine learning-enhanced data assimilation framework (DNN-ESMDA) that uses a dense neural network as a proxy for a pore network simulator, achieving significant computational speedup (thousands of hours to seconds) for inferring CO2-brine drainage relative permeability.
166. Where Not to Learn: Prior-Aligned Training with Subset-based Attribution Constraints for Reliable Decision-Making
Core Problem: Ensuring AI models justify decisions with acceptable evidence rather than shortcut correlations, by aligning learned representations with human priors, which is challenging due to divergence between model and human perception.
Key Innovation: Proposes an attribution-based human prior alignment method that encodes human priors as input regions and penalizes reliance on off-prior evidence during training, consistently improving task accuracy and decision reasonability in image classification and MLLM-based GUI agent models.
167. MAU-GPT: Enhancing Multi-type Industrial Anomaly Understanding via Anomaly-aware and Generalist Experts Adaptation
Core Problem: Automating fine-grained product image analysis for industrial quality control, which is hindered by limited dataset coverage and poor model generalization across diverse and complex anomaly patterns.
Key Innovation: Introduces MAU-Set, a comprehensive dataset for multi-type industrial anomaly understanding, and MAU-GPT, a domain-adapted multimodal large model with a novel AMoE-LoRA mechanism that unifies anomaly-aware and generalist experts adaptation, outperforming prior state-of-the-art methods.
168. VLRS-Bench: A Vision-Language Reasoning Benchmark for Remote Sensing
Core Problem: Existing remote sensing (RS) benchmarks are heavily biased toward perception tasks, hindering the development of Multimodal Large Language Models (MLLMs) for cognitively demanding RS applications that require complex reasoning.
Key Innovation: Introduction of VLRS-Bench, the first benchmark exclusively dedicated to complex RS reasoning, structured across Cognition, Decision, and Prediction, comprising 2,000 question-answer pairs spanning 14 tasks and up to eight temporal phases, constructed with RS-specific priors and expert knowledge.
169. ShapBPT: Image Feature Attributions Using Data-Aware Binary Partition Trees
Core Problem: Existing hierarchical Shapley approaches for eXplainable AI (XCV) do not exploit the multiscale structure of image data, leading to slow convergence and weak alignment with actual morphological features, and no prior method leveraged data-aware hierarchies for Computer Vision tasks.
Key Innovation: ShapBPT, a novel data-aware XCV method based on the hierarchical Shapley formula, which assigns Shapley coefficients to a multiscale hierarchical structure tailored for images (Binary Partition Tree), ensuring feature attributions align with intrinsic image morphology and reducing computational overhead.
170. Interpreting Physics in Video World Models
Core Problem: It is unclear whether video-based models rely on factorized representations of physical variables or implicitly represent them in a task-specific, distributed manner to make physically accurate predictions.
Key Innovation: The first interpretability study directly examining physical representations inside large-scale video encoders, identifying a 'Physics Emergence Zone' where physical variables become accessible and demonstrating that models use distributed, high-dimensional population structures for encoding variables like motion direction.
171. U-Net Based Image Enhancement for Short-time Muon Scattering Tomography
Core Problem: Short-time Muon Scattering Tomography (MST) suffers from poor image quality due to limited muon flux, hindering its practical application.
Key Innovation: A U-Net-based framework trained on simulated MST data to enhance image quality, significantly improving SSIM and LPIPS for low-statistics MST images.
172. BONSAI: Bayesian Optimization with Natural Simplicity and Interpretability
Core Problem: Standard Bayesian optimization (BO) does not aim to minimize deviation from a carefully engineered default configuration, often pushing weakly relevant parameters to the boundary of the search space, making it difficult to distinguish important from spurious changes and increasing the burden of vetting recommendations.
Key Innovation: Introduction of BONSAI, a default-aware BO policy that prunes low-impact deviations from a default configuration while explicitly controlling the loss in acquisition value, compatible with various acquisition functions, theoretically bounded in regret, and empirically shown to substantially reduce non-default parameters while maintaining competitive optimization performance.
173. Beyond Pooling: Matching for Robust Generalization under Data Heterogeneity
Core Problem: Pooling heterogeneous datasets across domains is a common strategy in representation learning, but naive pooling can amplify distributional asymmetries and yield biased estimators, especially in settings requiring zero-shot generalization.
Key Innovation: A matching framework that selects samples relative to an adaptive centroid and iteratively refines the representation distribution, using double robustness and propensity score matching to filter out confounding domains, achieving more robust generalization than naive pooling or uniform subsampling under asymmetric meta-distributions, and demonstrating improvements in zero-shot medical anomaly detection.
174. Exactly Computing do-Shapley Values
Core Problem: Computing do-Shapley values in Structural Causal Models (SCM) generally requires evaluating exponentially many terms, making it computationally expensive.
Key Innovation: Reformulates do-Shapley values in terms of the irreducible sets of the underlying SCM, enabling exact computation in time linear in the number of irreducible sets. It also provides an estimator and reduces the identification burden by proving non-parametric identifiability requires only the identification of interventional effects for singleton coalitions.
175. The Double-Edged Sword of Data-Driven Super-Resolution: Adversarial Super-Resolution Models
Core Problem: Data-driven super-resolution (SR) methods, often used as preprocessing in imaging pipelines, introduce a new, unexplored attack surface where adversarial behavior can be embedded directly into SR model weights, leading to downstream misclassification without input perturbations or backdoor triggers.
Key Innovation: Presents AdvSR, a framework that demonstrates adversarial behavior can be embedded into SR model weights during training. By jointly optimizing for reconstruction quality and targeted adversarial outcomes, AdvSR produces models that appear benign but induce downstream misclassification with high attack success rates and minimal quality degradation.
176. Optimization of Precipitate Segmentation Through Linear Genetic Programming of Image Processing
Core Problem: Current analysis of additive manufactured niobium-based copper alloys relies on slow, manual hand annotation for precipitate detection in micrographs due to varying image quality, hindering alloy development iteration speed.
Key Innovation: Develops a filtering and segmentation algorithm optimized using linear genetic programming (LGP) with a domain-specific language for image processing, producing human-interpretable MATLAB code that achieves near-human accuracy (1.8% error) and significantly faster processing (2 seconds per 3.6 MP image).
177. Row-Column Separated Attention Based Low-Light Image/Video Enhancement
Core Problem: U-Net based low-light image/video enhancement often results in local noise and loss of detail due to insufficient global information guidance, while traditional attention mechanisms significantly increase parameters and computations.
Key Innovation: Proposes a Row-Column Separated Attention module (RCSA) inserted after an improved U-Net, which utilizes global information from row and column feature map statistics to guide local information with fewer parameters, and introduces two temporal loss functions for video enhancement.
178. Data-Aware and Scalable Sensitivity Analysis for Decision Tree Ensembles
Core Problem: Existing sensitivity analysis methods for decision tree ensembles often produce unrealistic examples of sensitivity, limiting interpretability and practical value, and lack scalability for complex models.
Key Innovation: Proposes a data-aware sensitivity framework that constrains sensitive examples to be close to the training data, using novel MILP and SMT techniques, leading to more realistic and interpretable evidence of model weaknesses and improved scalability for large ensembles.
179. MUFASA: A Multi-Layer Framework for Slot Attention
Core Problem: Existing slot attention methods for unsupervised object-centric learning only utilize the last layer of pre-trained Vision Transformers (ViT), neglecting valuable semantic information encoded in other layers.
Key Innovation: Introduces MUFASA, a lightweight plug-and-play framework that computes slot attention across multiple feature layers of the ViT encoder and uses a fusion strategy to aggregate these multi-layer slots, improving segmentation results and training convergence for OCL methods.
180. Analyzing and Guiding Zero-Shot Posterior Sampling in Diffusion Models
Core Problem: Zero-shot diffusion-based methods for inverse problems (signal recovery) often rely on manual tuning and heuristics for incorporating observations, lacking a rigorous analysis and principled framework for parameter design.
Key Innovation: Provides a rigorous analysis of approximate posterior-samplers in diffusion models under a Gaussianity assumption, expressing them in closed-form in the spectral domain, and introduces a principled, method-agnostic framework for parameter design that yields tailored choices for each algorithm.
181. CausalTAD: Injecting Causal Knowledge into Large Language Models for Tabular Anomaly Detection
Core Problem: Existing LLM-based tabular anomaly detection methods randomly order columns without considering causal relationships, hindering accurate anomaly detection.
Key Innovation: CausalTaD injects causal knowledge into LLMs for tabular anomaly detection by identifying and reordering columns based on causal relationships and applying a reweighting strategy to enhance their effect.
182. VideoTemp-o3: Harmonizing Temporal Grounding and Video Understanding in Agentic Thinking-with-Videos
Core Problem: Conventional uniform frame sampling in long-video understanding fails to capture key visual evidence, leading to degraded performance and inefficiencies in agentic thinking-with-videos paradigms.
Key Innovation: VideoTemp-o3 is a unified agentic thinking-with-videos framework that jointly models video grounding and question answering, offering strong localization, on-demand clipping, and refinement of inaccurate localizations, improving long video understanding.
183. Open-Text Aerial Detection: A Unified Framework For Aerial Visual Grounding And Detection
Core Problem: Existing Open-Vocabulary Aerial Detection (OVAD) and Remote Sensing Visual Grounding (RSVG) methods are limited to coarse category-level semantics or single-target localization, preventing simultaneous rich semantic understanding and multi-target detection in aerial imagery.
Key Innovation: OTA-Det, a unified framework that bridges OVAD and RSVG paradigms through task reformulation, dense semantic alignment, and an efficient RT-DETR-based architecture, achieving state-of-the-art performance in open-text aerial detection with real-time inference.
184. Scalable Adaptation of 3D Geometric Foundation Models via Weak Supervision from Internet Video
Core Problem: Progress in 3D geometric foundation models is constrained by the scarcity of diverse, large-scale 3D annotations. Utilizing abundant Internet videos for scaling is challenging due to the absence of ground-truth geometry and the presence of observational noise.
Key Innovation: Introduction of SAGE, a framework for Scalable Adaptation of GEometric foundation models from raw video streams. It leverages a hierarchical mining pipeline and hybrid supervision (Sparse Geometric Anchoring via SfM point clouds and Dense Differentiable Consistency via 3D Gaussian rendering) to significantly enhance zero-shot generalization in 3D reconstruction.
185. Efficient Distribution Learning with Error Bounds in Wasserstein Distance
Core Problem: Learning an unknown probability distribution from finite samples with efficient, non-asymptotic, and easy-to-compute error bounds in Wasserstein distance is a fundamental challenge across various fields.
Key Innovation: A novel algorithmic and theoretical framework that approximates an unknown probability distribution by an approximate discrete distribution, bounding the Wasserstein distance between them with high confidence by solving a tractable optimization problem, and developing intelligent clustering algorithms to optimally find the support of the approximating distribution.
186. ViT-5: Vision Transformers for The Mid-2020s
Core Problem: Modernizing Vision Transformer backbones by systematically leveraging architectural advancements from the past five years to improve performance across various vision tasks.
Key Innovation: ViT-5, a new generation of Vision Transformers formed by component-wise refinement of normalization, activation functions, positional encoding, gating mechanisms, and learnable tokens, consistently outperforming state-of-the-art plain Vision Transformers on both understanding and generation benchmarks, and serving as a stronger backbone for generative modeling.
187. Variance-Gated Ensembles: An Epistemic-Aware Framework for Uncertainty Estimation
Core Problem: Existing methods for additively decomposing uncertainty in machine learning models often break down with finite-ensemble sampling or mismatched predictive distributions, leading to unreliable per-sample uncertainty estimation.
Key Innovation: Introduces Variance-Gated Ensembles (VGE), a differentiable framework that injects epistemic sensitivity via a signal-to-noise gate, providing a Variance-Gated Margin Uncertainty (VGMU) score and a Variance-Gated Normalization (VGN) layer for efficient and scalable epistemic-aware uncertainty estimation.
188. A Unified Framework for Multimodal Image Reconstruction and Synthesis using Denoising Diffusion Models
Core Problem: Existing methods for multimodal image reconstruction and synthesis require various task-specific models, complicating training and deployment workflows for handling incomplete multimodal imaging data.
Key Innovation: Any2all, a unified framework that formulates multimodal reconstruction and synthesis as a single virtual inpainting problem, using a single unconditional diffusion model trained on complete multimodal data, which can then 'inpaint' target modalities from any combination of inputs at inference time.
189. Fast Flow Matching based Conditional Independence Tests for Causal Discovery
Core Problem: Constraint-based causal discovery methods are computationally expensive due to the large number of conditional independence (CI) tests required, limiting their practical applicability.
Key Innovation: FMCIT (Flow Matching-based Conditional Independence Test), which leverages flow matching for high computational efficiency and requires only one-time model training, substantially accelerating causal discovery. It's integrated into GPC-FMCIT for improved accuracy-efficiency trade-offs.
190. UrbanGraphEmbeddings: Learning and Evaluating Spatially Grounded Multimodal Embeddings for Urban Science
Core Problem: Learning transferable multimodal embeddings for urban environments is challenging because urban understanding is inherently spatial, yet existing datasets and benchmarks lack explicit alignment between street-view images and urban structure.
Key Innovation: Introducing UGData, a spatially grounded dataset that anchors street-view images to spatial graphs with graph-aligned supervision; proposing UGE, a two-stage training strategy for progressively aligning images, text, and spatial structures; and creating UGBench, a comprehensive benchmark for evaluating spatially grounded embeddings in diverse urban understanding tasks.
191. Drop the mask! GAMM-A Taxonomy for Graph Attributes Missing Mechanisms
Core Problem: Understanding and addressing missing data in attributed graphs presents unique challenges beyond tabular datasets, as conventional missing data taxonomies do not account for graph-specific dependencies, leading to poor performance of existing imputation methods.
Key Innovation: Proposes GAMM (Graph Attributes Missing Mechanisms), a novel taxonomy that systematically links missingness probability to both node attributes and the underlying graph structure, demonstrating that state-of-the-art imputation methods struggle with these more realistic graph-aware missingness scenarios.
192. Is Meta-Path Attention an Explanation? Evidence of Alignment and Decoupling in Heterogeneous GNNs
Core Problem: It is unclear when meta-path attention in heterogeneous graph neural networks (GNNs) accurately reflects meta-path importance, and existing GNN explainers are not designed for heterogeneous graphs, confounding analysis.
Key Innovation: Introduction of MetaXplain, a meta-path-aware post-hoc explanation protocol for heterogeneous GNNs, and Meta-Path Attention--Explanation Alignment (MP-AEA) to empirically study and quantify the reliability of meta-path attention, revealing both alignment and decoupling regimes.
193. Bridging Academia and Industry: A Comprehensive Benchmark for Attributed Graph Clustering
Core Problem: A significant gap exists between academic research and real-world deployment of Attributed Graph Clustering (AGC) methods, due to limitations in current evaluation protocols such as small-scale datasets, non-scalable training, and reliance on supervised metrics.
Key Innovation: Presentation of PyAGC, a comprehensive, production-ready benchmark and library for AGC, featuring 12 diverse datasets (including industrial graphs), memory-efficient mini-batch implementations, and a holistic evaluation protocol emphasizing unsupervised structural metrics and efficiency.
194. FLAG-4D: Flow-Guided Local-Global Dual-Deformation Model for 4D Reconstruction
Core Problem: Existing methods for generating novel views of dynamic scenes struggle to capture complex point motions and fine-grained dynamic details consistently over time, especially from sparse input views.
Key Innovation: Introduces FLAG-4D, a novel framework that employs a dual-deformation network (Instantaneous Deformation Network and Global Motion Network) to dynamically warp 3D Gaussians, incorporating dense motion features from optical flow for higher-fidelity and temporally coherent 4D reconstructions of dynamic scenes.
195. Conditional Sequence Modeling for Safe Reinforcement Learning
Core Problem: Most offline safe reinforcement learning methods are trained for a pre-specified cost threshold, limiting their generalization and zero-shot deployment flexibility across varying cost constraints in practical scenarios.
Key Innovation: RCDT, a Conditional Sequence Modeling (CSM)-based offline safe RL algorithm, integrates a Lagrangian-style cost penalty with an auto-adaptive penalty coefficient, reward-cost-aware trajectory reweighting, and Q-value regularization to enable zero-shot deployment across multiple cost thresholds within a single trained policy, improving return-cost trade-offs.
196. Breaking the Grid: Distance-Guided Reinforcement Learning in Large Discrete and Hybrid Action Spaces
Core Problem: Standard Reinforcement Learning algorithms struggle with the curse of dimensionality in large discrete and hybrid action spaces, as existing methods rely on restrictive grid-based structures or computationally expensive nearest-neighbor searches.
Key Innovation: Distance-Guided Reinforcement Learning (DGRL) combines Sampled Dynamic Neighborhoods (SDN) for stochastic volumetric exploration in a semantic embedding space and Distance-Based Updates (DBU) to transform policy optimization into a stable regression task. This enables efficient RL in very large discrete and hybrid action spaces, improving performance and convergence speed.
197. The Theory and Practice of MAP Inference over Non-Convex Constraints
Core Problem: Efficiently and reliably computing constrained maximum a posteriori (MAP) predictions in probabilistic ML systems when subject to non-convex algebraic constraints is extremely challenging.
Key Innovation: Investigates conditions for exact and efficient constrained MAP inference and devises a general strategy that interleaves domain partitioning into convex feasible regions with numerical constrained optimization, outperforming constraint-agnostic baselines.
198. Low-Light Video Enhancement with An Effective Spatial-Temporal Decomposition Paradigm
Core Problem: Restoring dynamic or static scenes in low-light conditions, which are plagued by severe invisibility and noise, to achieve consistent and satisfactory enhancement.
Key Innovation: Proposing VLLVE and VLLVE++, innovative video decomposition strategies that incorporate view-independent and view-dependent components, a dual-structure enhancement network with cross-frame interaction, and an additive residual term for scene-adaptive degradations, enabling robust low-light video enhancement.
199. Rotated Lights for Consistent and Efficient 2D Gaussians Inverse Rendering
Core Problem: The high ambiguity in inverse rendering, particularly in estimating material and lighting from observed scene radiance, leading to inaccurate color and baked shadows in albedo estimation.
Key Innovation: Proposing RotLight, a simple capturing setup involving object rotation under varying light conditions to reduce ambiguity in inverse rendering, and introducing a proxy mesh for accurate incident light tracing and improved global illumination handling, leading to superior albedo estimation.
200. Closing the Confusion Loop: CLIP-Guided Alignment for Source-Free Domain Adaptation
Core Problem: Source-Free Domain Adaptation (SFDA) methods struggle in fine-grained scenarios due to asymmetric and dynamic class confusion, leading to noisy pseudo-labels and poor target discrimination.
Key Innovation: CLIP-Guided Alignment (CGA), a novel framework that explicitly models and mitigates class confusion in SFDA through three components: MCA for detecting confusion pairs, MCC for CLIP-guided confusion-aware pseudo-labeling, and FAM for aligning confusion-guided feature banks, leading to improved performance in confusion-prone and fine-grained scenarios.
201. From Correspondence to Actions: Human-Like Multi-Image Spatial Reasoning in Multi-modal Large Language Models
Core Problem: Multi-image spatial reasoning in MLLMs is challenging, as existing methods only partially or implicitly incorporate human-like mechanisms of cross-view correspondence and stepwise viewpoint transformation.
Key Innovation: HATCH, a training framework for MLLMs that explicitly incorporates human-aware mechanisms through Patch-Level Spatial Alignment and Action-then-Answer Reasoning, leading to improved multi-image spatial reasoning capabilities.
202. A Graphop Analysis of Graph Neural Networks on Sparse Graphs: Generalization and Universal Approximation
Core Problem: Existing analyses of Message Passing Graph Neural Network (MPNN) generalization and approximation capabilities are either only appropriate for dense graphs of unbounded sizes or for sparse graphs of uniformly bounded size, lacking a unified approach for graphs of all sizes.
Key Innovation: Presents a unified approach defining a compact metric on the space of graphs of all sizes (sparse and dense) under which MPNNs are H"older continuous, leading to more powerful universal approximation theorems and generalization bounds based on an extended graphop analysis.
203. Rethinking Graph Generalization through the Lens of Sharpness-Aware Minimization
Core Problem: Graph Neural Networks (GNNs) are highly sensitive to distribution shifts, leading to Minimal Shift Flip (MSF) where test samples slightly deviating from the training distribution are abruptly misclassified.
Key Innovation: Introduces Local Robust Radius to quantify loss sharpness, develops an energy-based formulation monotonically correlated with the robust radius, and proposes an energy-driven generative augmentation framework (E2A) to generate pseudo-OOD samples and enhance graph OOD generalization.
204. MotionCrafter: Dense Geometry and Motion Reconstruction with a 4D VAE
Core Problem: Reconstructing accurate 4D geometry and dense motion from monocular video is challenging, especially with prior methods that struggle with aligning 3D values and latents with RGB VAE latents due to fundamentally different distributions.
Key Innovation: Introduces MotionCrafter, a video diffusion-based framework that jointly reconstructs 4D geometry and estimates dense motion using a novel joint representation of dense 3D point maps and 3D scene flows in a shared coordinate system, and a 4D VAE. It employs a new data normalization and VAE training strategy to improve reconstruction quality, achieving state-of-the-art performance.
205. WorldCompass: Reinforcement Learning for Long-Horizon World Models
Core Problem: Long-horizon, interactive video-based world models struggle with accurate and consistent exploration based on interaction signals.
Key Innovation: Presents WorldCompass, an RL post-training framework that uses a clip-level rollout strategy, complementary reward functions (interaction-following accuracy and visual quality), and an efficient RL algorithm to significantly improve interaction accuracy and visual fidelity in world models.
206. When Simultaneous Localization and Mapping Meets Wireless Communications: A Survey
Core Problem: The paper surveys the intersection of SLAM and Wireless Communications, highlighting the need for robust joint solutions to enable highly effective retrieval of autonomous robot states and address challenges like scale ambiguity in monocular V-SLAM.
Key Innovation: A comprehensive survey of the state-of-the-art in joint communications and SLAM, attributing the bidirectional impact of each. It provides an overview of key concepts, techniques, and future directions, observing that RF information can resolve scale ambiguity in monocular V-SLAM, and visual odometry can benefit wireless communications in 5G and beyond.
207. Theory of Space: Can Foundation Models Construct Spatial Beliefs through Active Exploration?
Core Problem: Multimodal foundation models struggle with active, self-directed exploration to acquire information under partial observability and construct, revise, and exploit a spatial belief from sequential observations.
Key Innovation: Proposing 'Theory of Space' and a benchmark to evaluate foundation models' ability to actively explore and build cognitive maps, identifying critical bottlenecks like the Active-Passive Gap, inefficiency, belief instability, and Belief Inertia.
208. BayesFlow 2.0: Multi-Backend Amortized Bayesian Inference in Python
Core Problem: Modern Bayesian inference methods are often slow, especially for complex models and large datasets, making them prohibitive for many applications.
Key Innovation: BayesFlow 2.0, a Python library for general-purpose amortized Bayesian inference, offering rapid inference of model-implied quantities, support for multiple deep learning backends, customization, and new capabilities for hyperparameter optimization, design optimization, and hierarchical modeling.
209. Fast and Robust Likelihood-Guided Diffusion Posterior Sampling with Amortized Variational Inference
Core Problem: Zero-shot diffusion posterior sampling for inverse problems is flexible but computationally expensive, while existing amortized diffusion approaches are fast but lack robustness to unseen degradations.
Key Innovation: An amortization strategy for diffusion posterior sampling that preserves explicit likelihood guidance by amortizing inner optimization problems in variational diffusion posterior sampling, improving efficiency for in-distribution degradations while maintaining robustness to unseen operators.
210. $\partial$CBDs: Differentiable Causal Block Diagrams
Core Problem: Modern cyber-physical systems (CPS) require modeling frameworks that are simultaneously composable, learnable, and verifiable, but existing approaches treat these goals in isolation, lacking differentiability for learning in causal block diagrams or sufficient correctness guarantees in differentiable programming.
Key Innovation: Introduces differentiable causal block diagrams (∂CBDs), a unifying formalism that retains CBDs' compositional structure, incorporates assume-guarantee contracts for modular correctness, and introduces residual-based contracts as differentiable, trajectory-level certificates compatible with automatic differentiation, enabling a scalable, verifiable, and trainable modeling pipeline for CPS.
211. Learning-guided Kansa collocation for forward and inverse PDEs beyond linearity
Core Problem: Numerical methods for Partial Differential Equations (PDEs) suffer from the curse of dimensionality, high computation costs, and domain-specific discretization, limiting their application to complex scientific simulation problems.
Key Innovation: An extension of the CNF framework solver to multi-dependent-variable and non-linear settings for forward solutions, inverse problems, and equations discovery, including self-tuning techniques and evaluation on benchmark problems, aiming to overcome limitations of traditional PDE solvers.
212. Estimation of Fish Catch Using Sentinel-2, 3 and XGBoost-Kernel-Based Kernel Ridge Regression
Core Problem: Quantifying the connections between oceanographic factors and fish distribution to maintain fisheries and estimate fish catch using satellite imagery.
Key Innovation: Proposing an XGBoost-kernelized Kernel Ridge Regression (KRR) technique with Sentinel-2 and Sentinel-3 multispectral images for fish catch estimation, demonstrating improved capacity to capture nonlinear ocean-fish connections.
213. Learning the Value Systems of Societies with Preference-based Multi-objective Reinforcement Learning
Core Problem: Value-aware AI needs to recognize human values and adapt to diverse value systems, but operationalizing values is prone to misspecification, and their social nature requires representation across multiple users while accounting for group patterns.
Key Innovation: Proposes algorithms for jointly learning socially-derived value alignment models (groundings) and a set of value systems representing different user groups in a society, using clustering and preference-based multi-objective reinforcement learning (PbMORL) in MDPs.
214. Designing Multi-Robot Ground Video Sensemaking with Public Safety Professionals
Core Problem: Integrating multi-robot ground videos into public safety workflows for scalable situational awareness is challenging, and existing video sensemaking practices are burdensome.
Key Innovation: Developed a testbed for multi-robot ground video sensemaking with public safety-relevant events and a tool (MRVS) that augments video streams with a prompt-engineered video understanding model, reducing manual workload and increasing confidence.
215. Provably robust learning of regression neural networks using $\beta$-divergences
Core Problem: Regression neural networks trained with mean squared error are highly sensitive to outliers and data contamination, and existing robust methods often lack strong theoretical guarantees.
Key Innovation: Proposed 'rRNet', a new robust learning framework for regression NNs based on eta-divergence, providing provable theoretical guarantees on convergence, bounded influence functions, and an optimal 50% asymptotic breakdown point, addressing robustness to outliers and data contamination.
216. When do neural ordinary differential equations generalize on complex networks?
Core Problem: The generalization behavior of neural ordinary differential equations (neural ODEs) on graph-structured data, especially when applied to graphs with different sizes or structures than encountered during training, is poorly understood.
Key Innovation: Demonstrated that degree heterogeneity and the type of dynamical system are primary factors determining neural ODEs' ability to generalize across graph sizes and properties, and to capture fixed points and maintain performance amid missing data, providing insights into their application to complex systems.
217. Geometric Imbalance in Semi-Supervised Node Classification
Core Problem: Class imbalance in graph data, particularly in semi-supervised node classification, leads to 'geometric ambiguity' among minority-class nodes in the embedding space due to message passing.
Key Innovation: Formal introduction of 'geometric imbalance' and a unified framework to mitigate it through pseudo-label alignment, node reordering, and ambiguity filtering. The approach consistently outperforms existing methods under severe class imbalance.
218. View-Centric Multi-Object Tracking with Homographic Matching in Moving UAV
Core Problem: Multi-Object Tracking (MOT) in moving UAV scenarios is complex due to irregular flight trajectories, scene background changes, and significant view shifts, rendering traditional frame-to-frame object IoU association ineffective.
Key Innovation: HomView-MOT, a framework that harnesses view homography in changing scenes, incorporating a Fast Homography Estimation (FHE) algorithm and a Homographic Matching Filter (HMF) for object View-Centric ID Learning (VCIL) and more realistic physical IoU association.
219. Disentangled Representation Learning for Parametric Partial Differential Equations
Core Problem: Neural operators (NOs), while efficient PDE solvers, lack interpretability regarding underlying physical mechanisms due to their black-box nature and entangled representations of physical parameters.
Key Innovation: DisentangO, a novel hyper-neural operator architecture that learns disentangled representations of latent physical factors from NO parameters, enhancing physical interpretability and robust generalization across diverse PDE-governed systems.
220. Interpretable Generalized Additive Models for Datasets with Missing Values
Core Problem: Maintaining interpretability and sparsity in machine learning models is challenging when dealing with datasets containing missing values, as imputation or naive inclusion of indicator variables can complicate the model or sacrifice sparsity.
Key Innovation: Proposes M-GAM, a sparse, generalized additive modeling approach that incorporates missingness indicators and their interaction terms while maintaining sparsity through l0 regularization, achieving similar or superior accuracy with significantly improved sparsity.
221. Evaluating Sample Utility for Efficient Data Selection by Mimicking Model Weights
Core Problem: Large-scale web-crawled datasets often contain noise, bias, and irrelevant information, necessitating data selection techniques that are scalable and do not introduce unwanted data dependencies.
Key Innovation: Introduces the Mimic Score, a geometry-based data-quality metric that evaluates sample utility by measuring alignment between a sample's gradients and a target direction from a pre-trained reference model, and Grad-Mimic, a framework for online re-weighting and offline filtering to improve data efficiency and accelerate convergence.
222. Reducing Aleatoric and Epistemic Uncertainty through Multi-modal Data Acquisition
Core Problem: Modern AI systems struggle with disentangling and reducing aleatoric and epistemic uncertainty, especially when combining multi-modal data, impacting prediction accuracy and reliability.
Key Innovation: An innovative data acquisition framework that allows sampling in two directions (sample size and data modality) to reduce uncertainty, hypothesizing that aleatoric uncertainty decreases with more modalities and epistemic uncertainty with more observations, combining active learning, active feature acquisition, and uncertainty quantification.
223. A Survey on Class-Agnostic Counting: Advancements from Reference-Based to Open-World Text-Guided Approaches
Core Problem: Existing visual object counting methods are restricted to enumerating instances of known classes, requiring extensive labeled datasets and struggling in open-vocabulary settings, limiting their flexibility and generalizability.
Key Innovation: A comprehensive survey and taxonomy of Class-Agnostic Counting (CAC) methodologies, categorizing them into reference-based, reference-less, and open-world text-guided approaches, providing an overview of architectures, performance benchmarks, and future challenges.
224. CoHiRF: Hierarchical Consensus for Interpretable Clustering Beyond Scalability Limits
Core Problem: Existing clustering methods face computational and memory limits, struggle with high-dimensional noise, and lack stability, hindering their applicability to large-scale and complex datasets.
Key Innovation: CoHiRF, a hierarchical consensus framework that enables existing clustering methods to operate beyond scalability limits by repeatedly applying a base method to multiple feature views, enforcing agreement through consensus, and progressively reducing problem size, while also producing an interpretable Cluster Fusion Hierarchy.
225. LogicXGNN: Grounded Logical Rules for Explaining Graph Neural Networks
Core Problem: Existing rule-based explanations for Graph Neural Networks (GNNs) provide global interpretability but often optimize fidelity in an uninterpretable concept space, resulting in explanations that may appear faithful but are unreliable in their final subgraph form for end users.
Key Innovation: LogicXGNN, a post-hoc framework that constructs logical rules over reliable predicates explicitly designed to capture the GNN's message-passing structure, thereby ensuring effective grounding and improving data-grounded fidelity (Fid_D) by over 20% on average while being significantly faster.
226. Automatic and Structure-Aware Sparsification of Hybrid Neural ODEs
Core Problem: Hybrid neural ordinary differential equations (neural ODEs) suffer from training inefficiency and overfitting due to excessive latent states and interactions from mechanistic models, particularly in data-scarce settings.
Key Innovation: Proposes a new hybrid pipeline for automatic state selection and structure optimization in mechanistic neural ODEs. It combines domain-informed graph modifications with data-driven regularization to sparsify the model, improving predictive performance and stability while retaining mechanistic plausibility.
227. Rectified Flows for Fast Multiscale Fluid Flow Modeling
Core Problem: Statistical surrogate modeling of multiscale fluid flows is difficult due to their complex dynamics and sensitivity to initial conditions, and existing conditional diffusion surrogates require hundreds of stochastic sampling steps for inference.
Key Innovation: Proposes a rectified-flow surrogate that learns a time-dependent conditional velocity field, enabling fast, deterministic ODE-solve inference (8 steps vs. >=128 for diffusion models) while matching posterior statistics. It also provides a theoretical analysis of law-level conditional PDE forecasting and introduces a curvature-aware sampler for improved efficiency.
228. Revisiting Transformers with Insights from Image Filtering and Boosting
Core Problem: The self-attention mechanism in Transformers, despite its success, is largely heuristic-driven and lacks a deep mechanistic interpretation, hindering a robust theoretical understanding of its components and limitations.
Key Innovation: A unifying image processing framework that explains self-attention, positional encoding, and residual connections in Transformers, leading to two architectural modifications that empirically improve accuracy and robustness against data contamination and adversaries across language and vision tasks.
229. State-Space Hierarchical Compression with Gated Attention and Learnable Sampling for Hour-Long Video Understanding in Large Multimodal Models
Core Problem: Processing massive video-frame features from hour-long videos in large multimodal models leads to severe token explosion, making comprehensive video understanding computationally expensive and impractical.
Key Innovation: An efficient framework leveraging a bidirectional state-space model with a gated skip connection and learnable weighted-average pooling for hierarchical spatial and temporal downsampling, enabling cost-effective and competitive performance in hour-long video understanding while significantly reducing token budget.
230. These Are Not All the Features You Are Looking For: A Fundamental Bottleneck in Supervised Pretraining
Core Problem: Supervised pretraining often leads to deep learning models learning sparse feature representations that do not fully cover the needs of unseen downstream data, inducing transfer bias and limiting generalization.
Key Innovation: A theoretical framework demonstrating that pretraining captures inconsistent aspects of data distribution, and an inexpensive ensembling strategy that aggregates multiple models to generate richer feature representations, yielding a 9% improvement in transfer accuracy without extra pretraining cost.
231. Predicting Graph Structure via Adapted Flux Balance Analysis
Core Problem: Existing approaches for graph prediction in dynamic processes have limitations, such as assuming that vertices do not change between consecutive graphs, hindering accurate modeling of evolving network structures.
Key Innovation: An approach that predicts graph structure by exploiting time series prediction methods combined with an adapted form of flux balance analysis (FBA), incorporating various constraints applicable to growing graphs and demonstrating efficacy on synthetic and real datasets.
232. "PhyWorldBench": A Comprehensive Evaluation of Physical Realism in Text-to-Video Models
Core Problem: Video generation models have made progress in photorealism but struggle to accurately simulate physical phenomena, and there is a lack of comprehensive benchmarks to evaluate their adherence to the laws of physics.
Key Innovation: PhyWorldBench, a comprehensive benchmark designed to evaluate video generation models based on their adherence to physical laws across multiple levels of phenomena (object motion, energy conservation, rigid body interactions), including an 'Anti-Physics' category and a zero-shot evaluation method using multimodal large language models.
233. Mamba-based Spatio-Frequency Motion Perception for Video Camouflaged Object Detection
Core Problem: Existing video camouflaged object detection (VCOD) methods primarily rely on spatial appearances, which have limited discriminability due to high foreground-background similarity, and struggle with efficient long-sequence motion perception.
Key Innovation: Vcamba, a visual camouflage Mamba based on spatio-frequency motion perception, which integrates frequency and spatial features for efficient and accurate VCOD. It introduces a frequency-domain sequential scanning (FSS) strategy, an adaptive frequency enhancement (AFE) module, and space/frequency-based long-range motion perception modules (SLMP/FLMP) to model spatio-temporal and frequency-temporal sequences.
234. SNAP-UQ: Self-supervised Next-Activation Prediction for Single-Pass Uncertainty in TinyML
Core Problem: The absence of reliable and resource-efficient uncertainty estimation methods for on-device monitoring in TinyML, where microcontrollers must detect failures or distribution shifts under strict flash/latency budgets.
Key Innovation: SNAP-UQ, a novel and practical method for single-pass, label-free uncertainty estimation in TinyML based on depth-wise next-activation prediction, which is resource-efficient (smaller footprint, faster) and effective in detecting accuracy drops and failures in a single forward pass.
235. DyMixOp: A Neural Operator Designed from a Complex Dynamics Perspective with Local-Global Mixing for Solving PDEs
Core Problem: The primary challenge in using neural networks to accurately approximate nonlinear dynamical systems governed by partial differential equations (PDEs), especially when dynamics are non-linearizable or require infinite-dimensional spaces.
Key Innovation: DyMixOp, a novel neural operator framework for PDEs that integrates theoretical insights from complex dynamical systems, using a local-global mixing (LGM) transformation to effectively capture high-frequency details and complex nonlinear couplings, achieving state-of-the-art performance across diverse benchmark PDE systems.
236. Vision-Centric 4D Occupancy Forecasting and Planning via Implicit Residual World Models
Core Problem: Existing vision-centric world models for autonomous driving are inefficient as they fully reconstruct future scenes, expending significant capacity on redundantly modeling static backgrounds.
Key Innovation: IR-WM (Implicit Residual World Model), which focuses on modeling only the 'residual' changes in the environment by leveraging previous BEV features as a temporal prior. It also includes an alignment module to calibrate misalignments and demonstrates improved planning accuracy through implicit future state generation.
237. Decoupled Complementary Spectral-Spatial Learning for Background Representation Enhancement in Hyperspectral Anomaly Detection
Core Problem: Existing hyperspectral anomaly detection methods can be improved by enhancing background representation for better efficiency and robustness, especially without per-scene retraining.
Key Innovation: A decoupled complementary spectral-spatial learning framework that uses a two-stage training strategy (spectral enhancement via reverse distillation, then spatial branch as a complementary student) to jointly enhance background representations. This allows for deployment with parameter-free, training-free detectors.
238. How to Purchase Labels? A Cost-Effective Approach Using Active Learning Markets
Core Problem: Optimizing the acquisition of labeled data in resource-constrained environments to improve model performance for predictive analytics applications.
Key Innovation: Introducing and analyzing active learning markets with integrated budget constraints and improvement thresholds, using variance-based and query-by-committee strategies, which consistently achieve superior model performance with fewer labels acquired compared to conventional methods.
239. Free-RBF-KAN: Kolmogorov-Arnold Networks with Adaptive Radial Basis Functions for Efficient Function Learning
Core Problem: Original B-spline KANs suffer from high computational overhead, and existing RBF-based variants sacrifice approximation accuracy, creating a gap in efficient yet accurate function learning.
Key Innovation: Free-RBF-KAN, an architecture integrating adaptive learning grids and trainable smoothness parameters with RBFs, providing universal approximation proof for RBF-KANs and achieving comparable accuracy to B-spline KANs with significantly faster training and inference.
240. On Evaluation of Unsupervised Feature Selection for Pattern Classification
Core Problem: The evaluation of unsupervised feature selection methods using single-label datasets is unreliable and can lead to inconsistent performance rankings, as the chosen label arbitrarily influences perceived superiority.
Key Innovation: A multi-label classification framework for evaluating unsupervised feature selection methods, demonstrating that performance rankings differ significantly from single-label settings, thus providing a more fair and reliable comparison.
241. Federated Balanced Learning
Core Problem: In federated learning, non-iid data settings lead to client drift, negatively impacting global model performance, and existing methods primarily correct the deviated global model rather than addressing the root cause in client samples.
Key Innovation: Proposes Federated Balanced Learning (FBL) to prevent client drift by achieving sample balance on the client side through knowledge filling and sampling using edge-side generation models, complemented by a Knowledge Alignment Strategy and a Knowledge Drop Strategy.
242. Rethinking Cross-Modal Fine-Tuning: Optimizing the Interaction between Feature Alignment and Target Fitting
Core Problem: Adapting pre-trained models to unseen feature modalities for knowledge transfer is challenging, particularly in optimizing the interaction between feature alignment and target fine-tuning to avoid exacerbating misalignment and reducing generalization.
Key Innovation: Develops a principled framework with a provable generalization bound based on 'feature-label distortion' to explain and optimize the interaction between feature alignment and target fitting in cross-modal fine-tuning, leading to improved performance.
243. Visual Prompt-Agnostic Evolution
Core Problem: Existing Visual Prompt Tuning (VPT) variants for Vision Transformers suffer from unstable training dynamics, including gradient oscillations, early stagnation of shallow-layer prompts, and high-variance oscillations in deeper layers, leading to slow convergence and degraded performance.
Key Innovation: Prompt-Agnostic Evolution (PAE), a method that strengthens VPT by explicitly modeling prompt dynamics. It initializes prompts in a task-aware direction, uses a shared Koopman operator for coherent cross-layer evolution, and introduces a regularizer to constrain error amplification, leading to faster convergence and improved accuracy across various tasks and datasets.
244. PHAT: Modeling Period Heterogeneity for Multivariate Time Series Forecasting
Core Problem: Existing multivariate time series forecasting models largely neglect periodic heterogeneity, where variables exhibit distinct and dynamically changing periods, leading to suboptimal forecasting performance.
Key Innovation: Proposes PHAT (Period Heterogeneity-Aware Transformer) which effectively captures periodic heterogeneity by arranging inputs into a 'periodic bucket' tensor, restricting interactions within buckets, and employing a positive-negative attention mechanism with periodic priors, significantly outperforming existing methods.
245. Radial gradient of superionic hydrogen in Earth's inner core
Core Problem: Despite hydrogen being considered a major light element in Earth's core, the thermodynamics of its superionic phase and its distribution in the inner core remain unclear.
Key Innovation: Computes ab initio Gibbs free energies for liquid and superionic Fe-H phases to construct superionic-liquid phase diagrams. This reveals that solid-liquid partitioning is controlled primarily by a reduced temperature relative to iron melting and is weakly sensitive to pressure, leading to the discovery of a radial hydrogen gradient within the inner core based on thermochemical constraints.
246. When LLaVA Meets Objects: Token Composition for Vision-Language-Models
Core Problem: Current autoregressive Vision Language Models (VLMs) require a large number of visual tokens to represent images, leading to high computational costs, especially during inference.
Key Innovation: Introduction of Mask-LLaVA, a framework that leverages different levels of visual features (mask-based object, global, and local patch tokens) to create a compact and information-rich visual representation. This allows the model to flexibly reduce the number of mask-based object-tokens at test time without retraining, achieving competitive performance with fewer visual tokens.
247. Focus-Scan-Refine: From Human Visual Perception to Efficient Visual Token Pruning
Core Problem: Training-free token pruning methods for Vision-language models (VLMs) struggle to balance local evidence and global context under aggressive compression, leading to high inference latency and memory footprint.
Key Innovation: Proposes Focus-Scan-Refine (FSR), a human-inspired, plug-and-play pruning framework that first focuses on key evidence by combining visual importance with instruction relevance, then scans for complementary context, and finally refines the scanned context by aggregating nearby informative tokens. This consistently improves the accuracy-efficiency trade-off over existing state-of-the-art pruning methods for VLMs.
248. Vision Transformer Finetuning Benefits from Non-Smooth Components
Core Problem: The role of transformer architecture's smoothness in transfer learning, specifically finetuning performance, is poorly understood, and guidance is needed on which components to prioritize for adaptation.
Key Innovation: Demonstrates that high plasticity (low smoothness) of attention modules and feedforward layers consistently leads to better finetuning performance in vision transformers, offering a novel perspective that challenges the prevailing assumption that smoothness is always desirable.
249. Robustness Beyond Known Groups with Low-rank Adaptation
Core Problem: Deep learning models often exhibit systematic failures on unlabeled or unknown subpopulations, and existing group-robust methods typically assume prior knowledge of these subgroups.
Key Innovation: Proposes Low-rank Error Informed Adaptation (LEIA), a simple two-stage method that improves group robustness by identifying and adapting to a low-dimensional subspace where model errors concentrate, without requiring group labels or modifying the backbone.
250. A Global Optimization Algorithm for K-Center Clustering of One Billion Samples
Core Problem: The K-center clustering problem, which aims to minimize the maximum within-cluster distance, is challenging to solve globally optimally, especially for very large datasets.
Key Innovation: Develops a practical global optimization algorithm for K-center clustering based on a reduced-space branch and bound scheme, guaranteeing convergence to the global optimum and capable of handling one billion samples in parallel mode, significantly improving objective function values compared to heuristic methods.
251. Synthetic Oversampling: Theory and A Practical Approach Using LLMs to Address Data Imbalance
Core Problem: Imbalanced classification and spurious correlation are common challenges in machine learning, where underrepresented data samples compromise model accuracy, robustness, and generalizability. There is a lack of theoretical understanding for synthetic data approaches using Large Language Models (LLMs) to address these issues.
Key Innovation: Develops novel theoretical foundations to systematically study the roles of synthetic samples in addressing imbalanced classification and spurious correlation. It quantifies the benefits of synthetic oversampling, analyzes scaling dynamics, derives scaling laws, and demonstrates the capacity of transformer models to generate high-quality synthetic samples, validated through extensive numerical experiments.
252. On the Computational Efficiency of Bayesian Additive Regression Trees: An Asymptotic Analysis
Core Problem: The computational properties and convergence time of the widely-used Bayesian Additive Regression Trees (BART) sampler are not well-understood, particularly how its time to convergence increases with the number of training samples due to the multi-modal nature of the target posterior.
Key Innovation: Performs an asymptotic analysis of a modified BART sampler, showing that convergence time increases with training samples but can be dampened by increasing the number of trees or raising the sampler's temperature. These results provide a nuanced understanding of BART's computational efficiency and suggest strategies for improvement.
253. Q-Learning under Finite Model Uncertainty
Core Problem: Developing a robust Q-learning algorithm for Markov decision processes (MDPs) under model uncertainty, particularly when each state-action pair is associated with a finite ambiguity set of candidate transition kernels, which goes beyond common KL and Wasserstein ball formulations.
Key Innovation: Proposes a robust Q-learning algorithm that establishes almost sure convergence of the learned Q-function to the robust optimum and derives non-asymptotic high-probability error bounds. This finite-measure framework enables highly flexible, user-designed uncertainty models and can approximate Wasserstein ball and parametric ambiguity sets.
254. Fully Dynamic Adversarially Robust Correlation Clustering in Polylogarithmic Update Time
Core Problem: The challenge of maintaining an O(1)-approximation for dynamic correlation clustering with polylogarithmic update time, especially in an adversarially robust setting where edge labels can be adaptively flipped, has not been adequately addressed.
Key Innovation: Proposes a randomized algorithm that always maintains an O(1)-approximation to the optimal correlation clustering with O(log^2 n) amortized update time, even with adaptive edge label flips. This work significantly advances the state-of-the-art for adversarially robust dynamic correlation clustering and introduces a sparse-dense decomposition algorithm.
255. End to End Collaborative Synthetic Data Generation
Core Problem: Existing federated synthetic data generation techniques primarily focus on synthesizer training, neglecting crucial aspects like privacy-preserving preprocessing and evaluation, which are essential for an end-to-end collaborative framework for publishing synthetic data, especially for sensitive or rare datasets.
Key Innovation: Proposes an end-to-end collaborative framework for privacy-preserving synthetic data generation that accounts for both preprocessing and evaluation. This framework is instantiated with Secure Multiparty Computation (MPC) protocols and evaluated in a use case for privacy-preserving publishing of synthetic genomic data for leukemia.
256. Sparsified-Learning for High-Dimensional Heavy-Tailed Locally Stationary Time Series, Concentration and Oracle Inequalities
Core Problem: Developing a flexible sparse learning framework for high-dimensional heavy-tailed locally stationary time series, addressing challenges like smoothly changing regression functions and heavy-tailed noise.
Key Innovation: Introduction of a sparsity-inducing penalized estimation procedure combining additive modeling with kernel smoothing, derivation of concentration inequalities for kernel-weighted sums of locally stationary processes with heavy-tailed noise, and establishment of nonasymptotic prediction-error bounds.
257. Differentially Private Geodesic Regression
Core Problem: Extending classical linear regression to non-linear manifold spaces (geodesic regression) while ensuring differential privacy for the model parameters, especially when dealing with sensitive data.
Key Innovation: Proposing a Differentially Private (DP) K-Norm Gradient (KNG) mechanism for releasing parameters of geodesic regression on Riemannian manifolds, deriving theoretical sensitivity bounds tied to Jacobi fields and curvature, and demonstrating its efficacy on various manifold types.
258. Modern approaches to building interpretable models of the property market using machine learning on the base of mass cadastral valuation
Core Problem: Building effective and interpretable models for property markets using noisy real-world mass cadastral valuation data is challenging, especially when requiring interpretability for practical and legal reasons.
Key Innovation: Proposes a combination of classical linear regression with kriging for land parcels and the RuleFit method (linear regression with decision-tree-based rule generation) for flats, demonstrating that effective and interpretable property market models can be built, competitive with "black-box" methods like Random Forest.
259. Learning Geometric-Aware Quadrature Rules for Functional Minimization
Core Problem: The challenge of accurate numerical integration over non-uniform point clouds, which is critical for mesh-free machine learning solvers for PDEs using variational principles, as standard Monte Carlo methods are inadequate.
Key Innovation: QuadrANN, a Graph Neural Network (GNN) architecture that learns optimal quadrature weights directly from the underlying geometry of point clouds, generating a permutation-invariant and adaptive data-driven quadrature rule that reduces variance and improves stability for energy functional optimization.
260. Towards Spatio-Temporal Extrapolation of Phase-Field Simulations with Convolution-Only Neural Networks
Core Problem: Phase-field simulations of liquid metal dealloying (LMD) are computationally expensive, limiting their use for large domains and long time horizons.
Key Innovation: Introduces a fully convolutional, conditionally parameterized U-Net surrogate, coupled with a conditional diffusion model, to extrapolate LMD simulations far beyond training data in space and time, achieving significant speed-ups (up to 36,000x) while maintaining accuracy.
261. Task-free Adaptive Meta Black-box Optimization
Core Problem: Existing meta-black-box optimization (MetaBBO) methods require extensive handcrafted training tasks and predefined task distributions to learn meta-strategies, limiting their generalization to realistic applications with unknown task distributions.
Key Innovation: The Adaptive meta Black-box Optimization Model (ABOM) which performs online parameter adaptation using solely optimization data from the target task, introducing a closed-loop adaptive parameter learning mechanism for parameterized evolutionary operators, enabling zero-shot optimization.
262. Non-Uniform Noise-to-Signal Ratio in the REINFORCE Policy-Gradient Estimator
Core Problem: Policy-gradient methods in reinforcement learning often suffer from instability or slow training as learning progresses, due to an unclear understanding of the noise-to-signal ratio (NSR) of the REINFORCE estimator.
Key Innovation: Exact characterization of the noise-to-signal ratio (NSR) of the REINFORCE estimator for specific linear and polynomial systems, and a general upper bound for nonlinear dynamics, revealing that NSR is non-uniform and typically increases near an optimum, potentially causing training instability.
263. FlowSteer: Interactive Agentic Workflow Orchestration via End-to-End Reinforcement Learning
Core Problem: Existing agentic workflow orchestration faces key challenges including high manual cost, reliance on specific operators/LLMs, and sparse reward signals, limiting their efficiency and adaptability.
Key Innovation: FlowSteer, an end-to-end reinforcement learning framework that automates workflow orchestration through multi-turn interaction via a lightweight policy model and an executable canvas environment, supported by Canvas Workflow Relative Policy Optimization (CWRPO) for stable learning.
264. Near-Universal Multiplicative Updates for Nonnegative Einsum Factorization
Core Problem: Researchers lack user-friendly tools for fitting tailored nonnegative tensor factorizations to multiway data, often resorting to inefficient gradient-based methods or a limited set of mature implementations.
Key Innovation: NNEinFact, a near-universal einsum-based multiplicative update algorithm that can fit any nonnegative tensor factorization expressible as a tensor contraction by simply specifying the model with a string, supporting various loss functions and missing data, and demonstrating superior performance and speed compared to gradient-based methods.
265. Enhancing Mathematical Problem Solving in LLMs through Execution-Driven Reasoning Augmentation
Core Problem: Current multi-agent LLM systems for mathematical problem-solving lack reliably revisable reasoning processes, often operating in rigid sequential pipelines or relying on unreliable heuristic self-evaluation.
Key Innovation: Introduces Iteratively Improved Program Construction (IIPC), a reasoning method that iteratively refines programmatic reasoning chains by combining execution feedback with LLM's Chain-of-thought abilities, leading to improved mathematical reasoning performance.
266. ZKBoost: Zero-Knowledge Verifiable Training for XGBoost
Core Problem: There is a need for cryptographic guarantees of model integrity for machine learning models like XGBoost, especially in sensitive settings, to prove correct training without revealing data or parameters.
Key Innovation: Presents ZKBoost, the first zero-knowledge proof of training (zkPoT) protocol for XGBoost, which includes a fixed-point XGBoost implementation, a generic zkPoT template, and VOLE-based instantiation for proving nonlinear fixed-point operations.
267. Lyapunov Constrained Soft Actor-Critic (LC-SAC) using Koopman Operator Theory for Quadrotor Trajectory Tracking
Core Problem: Standard Reinforcement Learning (RL) algorithms lack stability guarantees, making them unsuitable for safety-critical physical systems where policies might induce oscillations or unbounded state divergence.
Key Innovation: Proposes a novel Lyapunov-constrained Soft Actor-Critic (LC-SAC) algorithm using Koopman operator theory to derive a closed-form Lyapunov function, which is then incorporated into SAC to provide stability guarantees for nonlinear systems, demonstrated on quadrotor trajectory tracking.
268. Beware Untrusted Simulators -- Reward-Free Backdoor Attacks in Reinforcement Learning
Core Problem: Untrusted simulators in Reinforcement Learning (RL) present a security blind spot, allowing adversarial developers to implant stealthy action-level backdoors into RL agents without altering or observing rewards.
Key Innovation: Proposes a novel attack called 'Daze' that exploits simulator dynamics to reliably and stealthily implant reward-free backdoors into RL agents, demonstrating its effectiveness and transferability to real robotic hardware.
269. Digital twin system for deepwater well construction: Enhancing operational efficiency and safety
Core Problem: Deepwater well construction faces challenges such as prolonged cycles, low efficiency, high emissions, and operational failures, influenced by operational and geotechnical factors affecting conductor installation and bearing capacity.
Key Innovation: Proposed an intelligent optimization algorithm integrating MT-FCNN, LSTM, and PSO, and established the first modularly coupled digital twin system for deepwater well construction, validated for predicting jetting flow rates, controlling ROP, and time-varying bearing capacity, providing a foundation for intelligent decision-making.
270. Influence of WEC geometric variables on the performance of an integrated WEC–VLFS system
Core Problem: The need for a solution that can both harvest wave energy and protect offshore structures, and understanding how WEC geometric variables influence the performance of such an integrated system.
Key Innovation: A novel integrated system combining a WEC array with a VLFS, modeled using the DMB method and Lagrange multiplier technique, demonstrating effective wave energy capture and significant reduction in VLFS hydroelastic response, particularly under short-wave conditions.
271. SSST-GAN: A Sampling-Based Spatial-Spectral Transformer and Generative Adversarial Network for Hyperspectral Unmixing
Core Problem: Existing transformer-based methods for hyperspectral unmixing struggle to effectively capture and fuse spatial-spectral features, account for nonlinear correlations, and are limited by predominant reliance on reconstruction error.
Key Innovation: SSST-GAN, a sampling-based spatial-spectral transformer and generative adversarial network that employs a dual-branch transformer encoder for independent spatial and spectral representation extraction, a feature enhancement module, a generalized nonlinear fluctuation model, and a generative adversarial learning framework to improve unmixing accuracy.
272. CRMF-Net: A Multimodal Fusion Network for Water–Land Classification From Single-Wavelength Bathymetric LiDAR
Core Problem: Accurate water-land classification from single-wavelength airborne LiDAR bathymetry (ALB) data is challenging due to the limited information content and feature ambiguity of one-dimensional waveform signals.
Key Innovation: CRMF-Net, a dual-branch multimodal fusion network that improves water-land classification accuracy and robustness by jointly exploiting complementary information from 1-D green waveform signals and their 2-D time-frequency representations, achieving high accuracy on CZMIL datasets.
273. Quantifying the spatial-seasonal patterns of land–atmosphere water, heat and CO2 flux exchange over the Tibetan Plateau from an observational perspective
Core Problem: Scarcity of land-atmosphere observation sites over the Tibetan Plateau, particularly in western and northern regions, limiting understanding of regional micro-climates, water cycles, energy budgets, and ecosystem dynamics.
Key Innovation: Establishment of a new research and observation platform with 16 planetary boundary layer towers on the Tibetan Plateau, providing a high-resolution, quality-controlled dataset for studying water-heat-carbon coupling and validating models.
274. 1 km annual forest cover and plant functional type dataset for China from 1981 to 2023
Core Problem: Existing land cover products often fail to capture the long-term increasing trend of forest area in China, leading to an underestimation of forest cover change-related ecological processes and impacts on the national terrestrial carbon balance.
Key Innovation: Development of a high-resolution (1 km), annual forest cover and plant functional type dataset for China (1981-2023) by integrating multisource remote sensing data with national forest inventories, improving simulations of biophysical and biogeochemical processes.
275. SoilHealthDB-V2: An updated and standardized global database of soil health under conservation management
Core Problem: Need for an updated and standardized global database of soil health indicators under conservation management to support more effective soil health assessments and related analyses.
Key Innovation: Development of SoilHealthDB-V2, an updated and standardized global database of soil health under conservation management, with increased observations, harmonized indicators, expanded spatiotemporal coverage, and new variables for detailed assessment of conservation practices and carbon cycling.
276. Hybrid feature-LSTM for solar radiation forecasting in different Chinese climate zones
Core Problem: Need for high-accuracy solar radiation forecasting to support agricultural development and national food security, which is challenging across different climate zones.
Key Innovation: Developed a high-accuracy solar radiation prediction system based on hybrid feature-LSTM models, incorporating feature importance analysis and optimization algorithms (specifically HLOA-LSTM with optimal meteorological inputs) to achieve superior forecasting precision in different Chinese climate zones.
277. Leveraging wide snapshot XCO<sub>2</sub> pre-training to estimate urban fossil fuel CO<sub>2</sub> emissions from space
Core Problem: Accurately identifying urban CO2 plumes and estimating emissions from satellite XCO2 snapshots, which is challenging due to broad spatial extent, low signal-to-noise ratio, and substantial data gaps.
Key Innovation: A Transformer-based deep learning model (leveraging masked pre-training on synthetic CO2M data) for XCO2 interpolation and plume detection, which improves gap-filling accuracy and significantly outperforms baselines in plume region segmentation, leading to improved consistency with bottom-up inventories for urban emission estimates from OCO-3 SAMs.
278. A hybrid physics-informed and data-driven model for estimating ocean internal wave phase speeds from remote sensing imagery
Core Problem: The scarcity of accurate and scalable datasets for ocean internal wave phase speeds, as traditional estimation methods rely on numerical simulations or sparse in-situ observations, limiting their accuracy and scalability.
Key Innovation: A hybrid physics-informed and data-driven model that integrates theoretical equations (KdV, BO, eKdV) as physical constraints, uses an adaptive ensemble learning framework to fuse data-driven and physical-informed features, and employs transfer learning to mitigate discrepancies, achieving high accuracy in estimating internal wave phase speeds from satellite imagery across varying water depths and globally.
279. Spatiotemporal dynamics of meteorological and groundwater droughts in Southwest China and their cumulative and lagged impacts on vegetation
Core Problem: Understanding the spatiotemporal evolution of meteorological and groundwater droughts in Southwest China and their cumulative and lagged impacts on vegetation, especially in complex karst hydrogeological conditions.
Key Innovation: Integrates multiple datasets (GRACE, GLDAS, SPEI, MODIS NDVI) to analyze drought dynamics and vegetation response, quantifying cumulative and lag effects in karst and non-karst areas, revealing that karst vegetation responds faster to groundwater drought and grasslands are more sensitive to meteorological drought.
280. Metaheuristic-optimized XGBoost model for accurate prediction of rock fragmentation in mining projects
Core Problem: Conventional methods for predicting rock fragmentation in mining projects are often inaccurate due to their inability to account for complex interactions between geology, blast design, and timing, leading to inconsistent cost estimations.
Key Innovation: Developed a metaheuristic-optimized XGBoost model (TTA_XGBoost) for accurate prediction of rock fragmentation, outperforming other optimized models, and identified key influencing features like joint angle and maximum charge per delay, providing a more reliable tool for mining project cost assessment.
281. Real-time multi-objective optimization and simulation of intelligent compaction for railway subgrade construction
Core Problem: Inefficient and quality-controlled compaction in railway subgrade construction, impacting subsequent stages and long-term operational safety.
Key Innovation: Proposes an ensemble Coupled Optimization–Evaluation Algorithm (COEA) integrating intelligent compaction, surrogate modeling, multi-objective optimization, and discrete event simulation for real-time decision-making, significantly enhancing compaction efficiency and reducing construction duration.