TerraMosaic Daily Digest: Feb 16, 2026
Daily Summary
This digest curates 240 papers spanning landslide deformation and runout, flood and coastal inundation, seismic hazards, and climate‑driven ground instability. A notable shift is from mapping to forecasting: SBAS‑InSAR time series aggregated to slope units are paired with ensemble learning (e.g., XGBoost) for long‑horizon displacement prediction, while multi‑temporal optical inventories track decade‑scale post‑event landscape recovery (e.g., Himalayan catchments after the 2013 Kedarnath disaster).
Across hazards, two threads are becoming operationally consequential. First, evidence‑grounded multimodal AI reframes impact assessment as synthesis, combining real‑time text signals with imagery to reduce peak‑extent blind spots (e.g., CrisiSense‑RAG; wildfire VLM pipelines). Second, hybrid physics–AI solvers and uncertainty‑aware surrogates accelerate coupled process simulation—from fluid–solid interaction frameworks for rapid mass movements to THM degradation in cold‑region tunnels and micro‑CT‑based characterization of freeze–thaw damage in landslide‑prone soils—without abandoning mechanism constraints needed for design and warning thresholds.
Key Trends
- Deformation monitoring moves toward prediction, not just detection: New workflows exploit time‑series InSAR at decision units (slope units, deformation “objects”), enabling spatially explicit forecasts and near‑real‑time updates even under rapid decorrelation (e.g., sequential MT‑PolInSAR).
- Multimodal crisis intelligence prioritizes evidence consistency over raw coverage: RAG and VLM systems increasingly fuse asynchronous sources—social reports for peak flooding, imagery for structural damage; faint‑signal satellite smoke detection plus language‑guided context—to produce auditable, time‑critical assessments.
- Physics‑informed surrogates scale coupled hazard simulation: Neural operators, PINNs, and diffusion‑based rollout control are being embedded in PDE pipelines, while coupled solvers (e.g., DDA–MPM for fluid–solid interaction) broaden the feasible scenario space for landslides, dam‑breaks, and coastal flooding.
- Cold‑region degradation is treated as a mechanistic driver, not a modifier: Freeze–thaw microstructure evolution (micro‑CT), cyclic stress‑path instability criteria in residual soils, and THM coupling in tunnels are tightening the link between climate forcing, material weakening, and failure thresholds.
- Risk and resilience research becomes more decision‑ready: Work on displacement and livelihood impacts (e.g., riverbank erosion), evacuation optimization, and census‑constrained exposure mapping signals a move from hazard intensity toward actionable consequence metrics under compound events.
Selected Papers
This digest features 240 selected papers from 1312 papers analyzed across multiple journals. Each paper has been evaluated for its relevance to landslide and broader geohazard research and includes links to the original publications.
1. Displacement prediction of slow-moving landslides using InSAR and ensemble regression models based on slope units
Core Problem: Accurately predicting displacement of slow-moving landslides is crucial for mitigation, but existing methods may not fully leverage InSAR data or provide spatially explicit long-term forecasts.
Key Innovation: Developed a novel predictive framework for slow-moving landslide displacements using SBAS InSAR time series aggregated at slope-unit scale and Extreme Gradient Boosting (XGBoost), achieving high accuracy (R² ≥ 0.98) for both trend and periodic terms, and providing spatially explicit long-term deformation forecasts for risk-based planning.
2. Instability mechanisms and collapse range prediction of steep high slopes during the open-pit to underground transition: A case study
Core Problem: The transition from open-pit to underground mining often induces instability in high, steep slopes, but existing stability analyses rarely consider the influence of underground mining-induced strata movement, hindering accurate instability mechanism elucidation and collapse range prediction.
Key Innovation: A new method for assessing slope stability and predicting surface subsidence range during open-pit to underground mining transition, based on a 2D mechanical analysis model, incorporating thrust reduction coefficient and subsidence stability index, and using a region-splitting iterative method, validated against simulations and field data.
3. Decadal evolution of mass movements in a Himalayan catchment after the catastrophic 2013 Kedarnath disaster
Core Problem: Understanding the decadal evolution and controlling factors of precipitation-triggered landslides and their recovery in mountainous terrain after extreme events, as less is known compared to earthquake-triggered landslides.
Key Innovation: Developed a multi-temporal landslide inventory using LISS-IV and PlanetScope satellite images (2014-2023) to track activity and recovery, identifying that activity was highest immediately after the event, declined, but renewed after 2018, and that areas near river confluences and tectonic thrusts remained hotspots.
4. WildfireVLM: AI-powered Analysis for Early Wildfire Detection and Risk Assessment Using Satellite Imagery
Core Problem: Wildfires are a growing threat, and early detection using satellite imagery is challenging due to faint smoke signals, dynamic weather, and the need for real-time analysis over large areas.
Key Innovation: WildfireVLM, an AI framework combining YOLOv12 for satellite imagery wildfire/smoke detection with MLLMs for language-driven contextualized risk assessments and response recommendations, deployed for real-time processing and tracking.
5. CrisiSense-RAG: Crisis Sensing Multimodal Retrieval-Augmented Generation for Rapid Disaster Impact Assessment
Core Problem: Automated disaster impact assessment struggles with temporal asynchrony between real-time human reports (capturing peak hazard conditions) and post-event satellite imagery (often reflecting recession), leading to dangerous underestimates of maximum extent.
Key Innovation: CrisiSense-RAG is a multimodal retrieval-augmented generation framework that reframes impact assessment as evidence synthesis. It employs hybrid dense-sparse retrieval for text and CLIP-based retrieval for imagery, with asynchronous fusion logic prioritizing real-time social evidence for peak flood extent and imagery for structural damage, achieving improved flood extent and damage severity MAE in zero-shot settings.
6. DeepC4: Deep Conditional Census-Constrained Clustering for Large-scale Multitask Spatial Disaggregation of Urban Morphology
Core Problem: Existing spatial disaggregation techniques for urban morphology mapping (e.g., building exposure, physical vulnerability) struggle with local discrepancies with census statistics and propagated model uncertainties, especially under weak and conditional label supervision.
Key Innovation: Introduces DeepC4, a novel deep learning-based spatial disaggregation approach that incorporates local census statistics as cluster-level constraints and considers multiple conditional label relationships in a joint multitask learning framework, enhancing the quality of urban morphology maps for building exposure and physical vulnerability.
7. Numerical modelling framework for assessing dune effectiveness against coastal inundation
Core Problem: Accurate coastal flood mapping and risk assessment are limited by gaps in data availability and model capabilities, particularly in simulating complex physical processes like wave setup, swash dynamics, and their interactions with protective infrastructure such as temporary dunes.
Key Innovation: An enhanced LISFLOOD-FP model that incorporates wave setup, swash dynamics, and interactions with temporary dunes, validated against real storm events, which provides a computationally efficient framework for assessing dune effectiveness against coastal inundation and supports practical risk assessment and infrastructure planning.
8. River bank erosion induced settlement displacement and livelihood impacts in Saghata Upazila of Gaibandha District
Core Problem: Riverbank erosion in Bangladesh causes significant settlement displacement and severe negative impacts on the livelihoods of affected populations, requiring quantification and assessment of these consequences.
Key Innovation: Quantified land erosion (20.84 km²) and settlement displacement (0.158 km²) in Saghata Upazila between 2000 and 2023 using satellite imagery and GIS, and assessed the severe decline in physical, human, financial, natural, and social capital-based livelihoods for displaced people.
9. Study on the microstructure of lateritic soil after freeze–thaw cycles under the background of climate warming
Core Problem: Lack of understanding of the microstructural changes in lateritic soil under freeze-thaw cycles, especially under climate warming conditions, which contribute to widespread landslides in regions like the Qinghai-Tibet Plateau.
Key Innovation: Quantitatively analyzed the microstructure (pore structure, porosity, pore size distribution) of lateritic soil after freeze-thaw cycles using X-ray micro-computed tomography (CT) under both fixed and elevated freezing temperatures, demonstrating that elevated temperatures significantly enhance pore connectivity and structural damage, providing insights for engineering design.
10. Deformation behaviors and potential instability criterion of granite residual soil: a perspective of cyclic mean effective stress under constant deviatoric stress path
Core Problem: Extreme rainfall events increase geological hazards in granite residual soil (GRS) regions, but the instability and deformation behaviors of GRS under the specific stress path of cyclic mean effective stress along a constant deviatoric stress path are not well understood.
Key Innovation: Conducted various triaxial tests to investigate GRS instability under cyclic loading, revealing a potential instability stress ratio and proposing a methodology utilizing the strain increment ratio-stress ratio curve from CSD tests to accurately predict instability under cyclic loading, providing critical references for geological hazard prevention.
11. Three-dimensional morphological characterization and quantitative roughness classification of loess joint surfaces
Core Problem: A detailed three-dimensional morphological characterization and quantitative roughness classification of loess joint surfaces, which act as fundamental structural controls on geological hazards like shear deformation and seepage-induced instability, was lacking.
Key Innovation: Comprehensive 3D morphological characterization and quantitative roughness classification of loess joint surfaces using multiple parameters and an entropy-weighted fuzzy comprehensive evaluation method, revealing anisotropy and the dominance of hydraulically driven processes in joint evolution, crucial for understanding loess geohazards.
12. Thermo-hydro-mechanical modeling for long-term performance estimation of cold-region tunnels
Core Problem: The spatiotemporal characteristics of freeze-thaw (F-T) damage in surrounding rock significantly impact the long-term performance and stability of cold-region tunnels, leading to complex thermo-hydro-mechanical (THM) evolution and potential failure.
Key Innovation: Proposed a novel THM coupling modeling framework that incorporates the spatiotemporal evolution of rock F-T degradation, including an F-T damage factor and dynamic parameter updating, to accurately assess surrounding rock deterioration and tunnel lining response under F-T cycling and climate warming.
13. A novel fluid–solid interaction framework by coupling three dimensional explicit discontinuous deformation analysis and material point method
Core Problem: Accurately capturing the complex dynamic interaction between fluids and arbitrarily shaped solids (FSI) remains a significant challenge in natural processes and engineering fields, particularly for disaster simulations.
Key Innovation: Proposed a novel 3D FSI framework coupling explicit Discontinuous Deformation Analysis (DDA) for solids and Material Point Method (MPM) for fluids, validated it with benchmarks including underwater landslides and dam break impacts, demonstrating its potential for complex FSI problems in geotechnical engineering.
14. Empirically Based Constitutive Modeling Approach for Methane Hydrate‐Bearing Soils
Core Problem: Existing constitutive models for methane hydrate-bearing sediments (MHBS) are often adapted from conventional soils and may not accurately capture their complex mechanical behavior, leading to uncertainties in simulating engineering processes like methane extraction and predicting stability.
Key Innovation: A new empirically-based approach to derive essential mechanical features (yield criterion, flow rule, strain-hardening) directly from experimental observations, independent of prior modeling assumptions, leading to a tailored constitutive model for MHBS that better captures their mechanical response.
15. An Efficient Global Automatic Threshold Detection Algorithm for Large‐Scale Flood Distribution Analysis
Core Problem: Identifying the optimal threshold for peaks over threshold (POT) series is crucial but challenging for effective flood distribution analysis and risk reduction, especially at large scales.
Key Innovation: Proposed an automatic threshold detection method based on the Shuffled Complex Evolution (SCE-UA) optimization algorithm, which efficiently locates the global optimal threshold without objective specification. Applied to 380 stations in China, it improved accuracy and provided insights into flood tail characteristics.
16. Foundation Model-Driven Semantic Change Detection in Remote Sensing Imagery
Core Problem: Existing semantic change detection (SCD) methods in remote sensing imagery face significant challenges in performance and paradigm complexity due to limited semantic understanding and the inherent complexity of SCD tasks.
Key Innovation: PerASCD is a foundation model-driven SCD method that enhances multi-scale semantic understanding and overall performance using the RS foundation model PerA. It introduces a modular Cascaded Gated Decoder (CG-Decoder) to simplify decoding and a Soft Semantic Consistency Loss (SSCLoss) to mitigate training instability, achieving state-of-the-art performance.
17. S2SServiceBench: A Multimodal Benchmark for Last-Mile S2S Climate Services
Core Problem: There is a 'last-mile gap' in subseasonal-to-seasonal (S2S) climate services: translating scientific forecasts into trusted, actionable climate services, requiring reliable multimodal understanding and decision-facing reasoning under uncertainty.
Key Innovation: Introduces S2SServiceBench, a multimodal benchmark curated from an operational climate-service system to evaluate MLLMs and agents in generating decision-making deliverables from S2S forecasts, covering 10 service products across six application domains including Disasters.
18. Event-based Visual Deformation Measurement
Core Problem: Traditional visual deformation measurement methods struggle with highly dynamic scenes or require high-speed cameras, leading to prohibitive storage and computational costs, limiting their applicability for dense deformation field recovery.
Key Innovation: Proposes an event-frame fusion framework that exploits events for temporally dense motion cues and frames for spatially dense precise estimation, combined with an Affine Invariant Simplicial (AIS) framework to mitigate motion ambiguities and a neighborhood-greedy optimization strategy for efficient parameter searching and error reduction.
19. Architectural Insights for Post-Tornado Damage Recognition
Core Problem: Current automated methods for post-tornado building damage assessment struggle with the unique visual complexity of tornado wreckage, severe domain shift from pre-training data, and extreme class imbalance, hindering rapid and accurate assessment.
Key Innovation: A systematic experimental framework evaluating 79 deep learning models on a new QSTD benchmark dataset, revealing that optimizer choice (e.g., SGD over Adam) and a low learning rate are more critical than architecture alone for achieving operational-grade performance in post-tornado damage recognition, leading to significant F1 gains and strong cross-event generalization.
20. DiffusionRollout: Uncertainty-Aware Rollout Planning in Long-Horizon PDE Solving
Core Problem: Autoregressive diffusion models for long-horizon predictions of physical systems governed by PDEs suffer from error accumulation, leading to reduced reliability over extended trajectories.
Key Innovation: Introduces DiffusionRollout, an uncertainty-aware selective rollout planning strategy that uses predictive uncertainty (standard deviations) to adaptively select step sizes, thereby mitigating error accumulation and improving long-term prediction reliability for PDEs.
21. GREAT-EER: Graph Edge Attention Network for Emergency Evacuation Responses
Core Problem: Developing effective and fast evacuation plans for urban areas during emergency situations (including natural disasters) is challenging due to the NP-hard nature of the Bus Evacuation Orienteering Problem (BEOP) and the need for rapid decision-making.
Key Innovation: Introduction of GREAT-EER, a deep reinforcement learning-based method utilizing graph learning to solve the BEOP, achieving fast inference speed and near-optimal evacuation plans for bus-based evacuation in urban areas, demonstrated on real-world scenarios.
22. Seismic event classification with a lightweight Fourier Neural Operator model
Core Problem: Rapid and accurate real-time classification of triggered data from continuous seismic data streams for induced seismicity monitoring, which is computationally intensive for existing deep learning models.
Key Innovation: Proposes a lightweight Fourier Neural Operator (FNO) model for microseismic event classification, demonstrating high effectiveness (F1 score 95-98%) with significantly reduced computational cost, making it suitable for resource-constrained real-time deployment.
23. Dynamic analysis for monopile offshore wind turbine under combined wind-wave-seismic loading: Effects of clay cyclic degradation and shakedown
Core Problem: Combined wind, wave, and seismic loads induce significant variations in marine clay properties (cyclic degradation/shakedown), and their influence on the dynamic response of monopile offshore wind turbines (OWTs) is not fully understood.
Key Innovation: Develops an enhanced bounding surface elastoplastic model for marine clay and a 3D finite element model of the soil-monopile-OWT system to evaluate the effects of cyclic clay behaviors, load combinations, and seismic intensity, finding that cyclic shakedown causes significant acceleration amplification and that small-strain models attenuate acceleration while amplifying displacement.
24. SSI effects on dynamic responses of a large-scale offshore wind turbine under earthquake loadings
Core Problem: Understanding the nonlinear soil-structure interaction (SSI) effects on the seismic responses of large-scale offshore wind turbines (OWTs) is crucial for their safe design, especially considering potential overestimations if SSI is neglected.
Key Innovation: Develops a novel coupled analysis framework based on OpenFAST, modified to incorporate SSI modeling and seismic analysis, to predict seismic responses of OWTs, demonstrating that neglecting SSI can significantly overestimate mudline bending moments (up to 126%) and that parked states may not be optimal during earthquakes.
25. An integrated water deficit index to evaluate multifaceted impacts of drought in a semi-arid river basin
Core Problem: Existing drought indices often fail to capture overall water availability by considering all storage fluxes, particularly for large river basins, making it difficult to evaluate the multifaceted impacts of drought.
Key Innovation: Developed an Integrated Water Deficit Index (IWDI) by combining standardized precipitation evapotranspiration index (SPEI) and terrestrial water storage index (TWSI) using Clayton copula, demonstrating its robustness in characterizing spatiotemporal drought variations and quantifying impacts like reduced water bodies (15%) and crop production (25%).
26. Rainfall relief or humidity havoc? The boon and curse of precipitation during heatwaves
Core Problem: The complex interaction between heat, humidity, and precipitation during heatwaves is not fully understood, particularly how precipitation can both mitigate heat and exacerbate health risks by increasing humidity.
Key Innovation: Analyzed historical data (1980-2020) across the contiguous US, revealing a noticeable increase in heatwave duration and humidity levels, and showing that precipitation events terminating heatwaves often lead to heightened humidity, compounding health risks and highlighting the need for integrated climate adaptation strategies.
27. Monitoring river restoration structures through velocimetry: a field-based study on the use of submerged vanes for debris mitigation along a bridge
Core Problem: Lack of field-based testing and understanding of hydrodynamic and geomorphological interactions induced by in-stream flow deflecting structures (submerged vanes) designed to mitigate debris accumulation and localized scour along bridge piers, which can lead to catastrophic bridge failure.
Key Innovation: Conducted a field-based study using Acoustic Doppler Current Profiler (ADCP) and Large-Scale Particle Image Velocimetry (LSPIV) to monitor the effectiveness of submerged vanes in mitigating debris and scour along a bridge, demonstrating their influence on sedimentary dynamics and flow patterns, though with reduced efficacy during high-flow events.
28. Seismic Fragility Analysis of Rock Tunnels Using Multi-Stripe Analysis
Core Problem: A deeper understanding of rock tunnel vulnerability to seismic events is needed, and current design guidelines may be insufficient for ensuring seismic resilience.
Key Innovation: Developed seismic fragility curves for shallow rock tunnels using a performance-based approach with 2D nonlinear time-history analyses and Multi-stripe Analysis, providing a detailed assessment of seismic vulnerability and highlighting limitations of current design guidelines.
29. Hybrid Soil Stabilisation Using Cement and Nanomaterials: A Systematic Review and Meta-Analysis
Core Problem: A unified and quantitative synthesis of hybrid cement-nanomaterial systems for soil stabilization, particularly concerning their combined application and mechanistic understanding, was lacking despite the shift from conventional methods.
Key Innovation: A comprehensive systematic review and meta-analysis of hybrid cement-nanomaterial soil stabilization, providing an integrated mechanistic synthesis, quantitative evaluation (7.20x strength enhancement), and a sustainability-driven framework for improving geotechnical properties and resistance to collapsibility and liquefaction.
30. Geopolymer Crusts from Waste Marble Powder for Sand Dune Stabilization in Arid Climates: A Multi-Performance Assessment
Core Problem: Wind-driven sand dune migration poses significant environmental and socio-economic challenges, necessitating sustainable and effective sand dune fixation techniques, especially those valorizing industrial waste.
Key Innovation: Development and multi-performance assessment of geopolymer-stabilized sand crusts from waste marble powder for sand dune fixation, demonstrating rapid strength gains, superior erosion resistance, and environmental safety, offering a cost-effective and eco-friendly solution to combat land degradation.
31. Investigating GIA-driven intraplate deformation in Eastern Canada with GNSS clustering and 3D strain rate analysis
Core Problem: Deciphering the mechanisms driving the spatiotemporal distribution of intraplate seismicity in Eastern Canada and understanding seismic risks associated with Glacial Isostatic Adjustment (GIA).
Key Innovation: Applies machine learning cluster analysis and 3D strain rate analysis to GNSS data, identifying zones of different deformation styles aligned with GIA signals. Quantifies volumetric strain rates that may be sufficient to reactivate ancient faults, contributing to understanding future seismicity.
32. Characterization of energy dissipation and modeling of damage evolution in frozen soils under cyclic impact loading
Core Problem: Frequent disturbances from impacts in cold regions contribute to soil failure and instability, but existing studies lack detailed characterization of energy dissipation and a comprehensive constitutive model for frozen soil under cyclic impact loading.
Key Innovation: Conducted impact experiments to characterize energy dissipation in frozen soil and developed a novel cyclic dynamic degradation model that considers defect evolution, thermal damage, and an elastic modulus reduction factor, making it applicable to repeated impact loading in cold regions.
33. A dual mortar method for analyzing the effects of wave-induced instantaneous liquefaction on an immersed tunnel with a liquefaction-associated non-Darcy flow model
Core Problem: Extreme ocean waves can cause instantaneous liquefaction of the seabed, leading to decreased bearing capacity and potential damage to subsea structures like immersed tunnels, and existing models may not adequately capture nonphysical tensile behavior or robustly treat seabed-tunnel interfacial conditions.
Key Innovation: Incorporated a liquefaction-associated non-Darcy flow model into a wave–seabed–tunnel model and developed a dual mortar method to numerically treat special interfacial conditions, demonstrating that extreme waves can pose significant influences on immersed tunnels through liquefaction.
34. LAF-YOLOv10 with Partial Convolution Backbone, Attention-Guided Feature Pyramid, Auxiliary P2 Head, and Wise-IoU Loss for Small Object Detection in Drone Aerial Imagery
Core Problem: Current object detectors struggle with UAV-specific challenges like small targets, cluttered backgrounds, occlusion, and computational budgets for small object detection in drone aerial imagery.
Key Innovation: Introducing LAF-YOLOv10, an enhanced YOLOv10n model integrating Partial Convolution C2f, Attention-Guided Feature Pyramid Network, an auxiliary P2 detection head, and Wise-IoU v3 loss, achieving improved small-object detection performance and viability for embedded UAV deployment in disaster response and other applications.
35. A Deep Convolutional Network to Extract Real-Time Landmarks for UAV Navigation
Core Problem: UAV navigation in GNSS-denied environments is challenging due to signal degradation or loss, making reliable landmark extraction from onboard camera images critical for positioning in monitoring applications.
Key Innovation: Proposes a convolution-based deep learning approach for real-time extraction of appropriate visual landmarks from UAV images, enabling navigation without GNSS support, which is critical for monitoring applications.
36. Ice-free geomorphometry of Queen Maud Land, East Antarctica: 3. Belgica and Yamato (Queen Fabiola) Mountains
Core Problem: There is a need for new quantitative knowledge about the topography of unique ice-free Antarctic landscapes and for morphometric information to support various Antarctic research fields.
Key Innovation: Performed geomorphometric modeling and mapping of the Belgica and Yamato Mountains using REMA data, deriving and presenting models and maps of eleven scientifically important morphometric variables (e.g., slope, curvature, topographic wetness index), providing rigorous, quantitative, and reproducible data for future geological, geomorphological, glaciological, ecological, and hydrological studies.
37. A unified framework for evaluating the robustness of machine-learning interpretability for prospect risking
Core Problem: Machine learning-based classifiers are used for hydrocarbon prospect risking, but their lack of transparency (black-box nature) and the disagreement or differences in explanations generated by various explainable AI (XAI) methods (e.g., LIME, SHAP) for the same scenario undermine the trustworthiness and reliability of these explanations, especially for complex data.
Key Innovation: Proposes a unified framework to generate counterfactuals and quantify necessity and sufficiency, using these to perform a robustness evaluation of explanations provided by LIME and SHAP on high-dimensional structured prospect risking data, thereby providing deeper insights into model capabilities and identifying optimal XAI-model pairings for hydrocarbon indication.
38. Locally Private Parametric Methods for Change-Point Detection
Core Problem: Identifying distributional changes in time series data (change-point detection) while ensuring local differential privacy, and characterizing the statistical cost of privacy on detection accuracy.
Key Innovation: Derives improved finite-sample accuracy guarantees for non-private change-point detection and proposes two locally differentially private algorithms, characterizing the statistical cost of privacy and demonstrating their performance in detecting changes in time series.
39. LEAD-Drift: Real-time and Explainable Intent Drift Detection by Learning a Data-Driven Risk Score
Core Problem: Conventional approaches for 'intent drift' detection in Intent-Based Networking (IBN) only raise alarms when degradation is significant, limiting their effectiveness for proactive failure prevention.
Key Innovation: Introduces LEAD-Drift, a framework that reformulates intent failure detection as a supervised learning problem to predict a future risk score in real-time, providing significantly earlier and less noisy warnings, multi-horizon time-to-failure estimation, and per-alert explainability using SHAP.
40. Benchmarking AI-based data assimilation to advance data-driven global weather forecasting
Core Problem: The rapid expansion of AI-based Data Assimilation (DA) research for weather forecasting lacks an objective, comprehensive, and real-world benchmark for fair comparison and evaluation of diverse methods.
Key Innovation: Introduction of DABench, a benchmark integrating real-world observations for developing and evaluating AI-based DA methods, demonstrating competitive performance with state-of-the-art AI-driven four-dimensional variational frameworks for global weather DA and medium-range forecasting.
41. MixLinear: Extreme Low Resource Multivariate Time Series Forecasting with 0.1K Parameters
Core Problem: Long-term Time Series Forecasting (LTSF) faces significant challenges due to complex temporal dependencies and high computational demands, making existing high-accuracy models unsuitable for resource-constrained devices.
Key Innovation: Introduction of MixLinear, an ultra-lightweight multivariate time series forecasting model (0.1K parameters) designed for resource-constrained devices, which effectively captures both temporal and frequency domain features, achieving comparable or superior performance to state-of-the-art models with significantly reduced computational overhead.
42. OmniEarth-Bench: Towards Holistic Evaluation of Earth's Six Spheres and Cross-Spheres Interactions with Multimodal Observational Earth Data
Core Problem: Existing multimodal learning benchmarks in Earth science offer limited, siloed coverage of Earth's spheres and their cross-sphere interactions, lacking broad data heterogeneity, scientific granularity, and extensibility, particularly for the lithosphere.
Key Innovation: Introduces OmniEarth-Bench, the first multimodal benchmark systematically spanning all six Earth spheres (including lithosphere) and cross-sphere interactions, built with a scalable data inference framework and 29,855 expert-curated annotations across 109 tasks. It reveals systematic gaps in current MLLMs' Earth-system cognitive ability.
43. Lightning Prediction under Uncertainty: DeepLight with Hazy Loss
Core Problem: Existing lightning prediction models struggle with capturing dynamic spatial context and inherent randomness, underutilize key observational data, and rely heavily on computationally expensive and sensitive Numerical Weather Prediction (NWP) systems.
Key Innovation: Presents DeepLight, a novel deep learning architecture leveraging multi-source meteorological data through a dual-encoder and multi-branch convolution, introducing a Hazy Loss function to explicitly address spatio-temporal uncertainty, significantly improving lightning prediction accuracy.
44. Coherent Source Subsampling: A Data-Driven Strategy for Restoring Causal-Acausal Symmetry in Ambient Seismic Wavefield Correlations
Core Problem: Ambient noise tomography produces biased estimates of Green's functions due to spatially and temporally heterogeneous ambient noise sources, violating the equipartition assumption and leading to inaccurate surface-wave dispersion measurements.
Key Innovation: Proposes Coherent Source Subsampling (CSS), a data-driven method that selects and averages only cross-correlation time windows associated with stationary zone source excitation, thereby mitigating non-uniform source distribution effects and restoring causal-acausal symmetry for more stable and accurate surface-wave dispersion measurements and tomograms.
45. Efficient Tensor Completion Algorithms for Highly Oscillatory Operators
Core Problem: Efficiently reconstructing highly oscillatory operators from limited observations, particularly in seismic applications, where existing methods are computationally expensive or less accurate.
Key Innovation: Presents low-complexity tensor completion algorithms based on butterfly decomposition for highly oscillatory operators, achieving $O(n \log^3 n)$ computational cost and significantly smaller reconstruction errors compared to state-of-the-art methods, demonstrated in seismic applications.
46. Study on anti-overturning characteristics and analysis methods of bucket foundations for offshore wind turbine considering translational effects
Core Problem: The overturning bearing behavior of novel bucket foundations for offshore wind turbines requires in-depth investigation for design and safety assessment, especially considering translational effects.
Key Innovation: Proposed a deformation-controlled analytical method for assessing the anti-overturning stability of bucket foundations, incorporating translational decomposition of foundation displacement, verified by FE numerical simulation, providing valuable references for structural optimization and safety assessment.
47. Hydrological regime shifts in Sahelian watersheds: an investigation with a simple dynamical model driven by annual precipitation
Core Problem: Understanding and identifying the timing of hydrological regime shifts in Sahelian watersheds, characterized by increased annual runoff coefficients during and after severe droughts, and the underlying soil-water-vegetation feedbacks.
Key Innovation: Introduction of a simple lumped dynamical model, driven by annual precipitation, to investigate hydrological regime shifts in Sahelian watersheds, demonstrating that the model can reproduce runoff coefficient trends, that most parameterizations are bistable, and identifying specific years of regime shifts for four watersheds.
48. Hotspots and hot moments of metal mobilization: dynamic connectivity in legacy mine waters
Core Problem: Conventional monitoring and treatment of contaminated mine water overlook the initial stages of pollutant release and their dynamic pathways within abandoned mines, leading to inefficient 'end-of-pipe' remediation strategies.
Key Innovation: The study reveals that episodic shifts in subsurface hydrological connectivity govern metal(loid) mobilization from localized storage zones in abandoned mines, identifying 'hotspots' and 'hot moments' of pollution and suggesting more effective source-related, decentralized treatment strategies.
49. In situ Temperature Monitoring in a Large Scale Artificial Ground Freezing Project
Core Problem: Comprehensive observation studies on large-scale artificial ground freezing (AGF) projects are scarce, making it difficult to understand temperature displacement, optimize freezing effects, and manage cooling energy demand efficiently.
Key Innovation: In situ temperature monitoring in a large-scale AGF project, revealing frozen curtain deterioration after excavation and proposing/verifying a new method for evaluating freezing effects and dynamic cooling regulation, providing a practical reference for similar large-scale projects.
50. Stress prediction equation for buried pipelines subjected to internal loading and cold wave-induced soil temperature
Core Problem: Buried water pipelines are vulnerable to failure under complex thermo-mechanical loading during extreme cold waves, but current analytical methods neglect thermo-mechanical coupling, limiting accurate stress estimation.
Key Innovation: Developed a novel closed-form predictive equation for maximum stress in buried pipelines, explicitly incorporating thermo-mechanical coupling effects through extensive 3D finite element analysis, providing a practical and accurate tool for assessing pipe performance under cold wave conditions.
51. Novel fuzzy-based evaluation of scalar- and vector-valued intensity measures in seismic fragility analysis of shallow-buried rectangular subway station structures
Core Problem: The need for reliable selection of optimal scalar- and vector-valued intensity measures (IMs) for accurate seismic fragility analysis and probabilistic seismic demand models (PSDM) of shallow-buried subway station structures.
Key Innovation: Proposes a novel fuzzy-based method for evaluating and selecting optimal SIMs and VIMs, including a new framework for VIM evaluation that considers contributions and correlations of internal IMs. It demonstrates that VIMs reduce estimation bias compared to SIMs.
52. Fluid flow through fractured rock masses: Insight from single to complex fracture networks
Core Problem: A comprehensive understanding of fluid flow through fractured rock masses, from single fractures to complex networks, including the effects of shear, is crucial for assessing hydraulic properties and engineering applications relevant to rock stability.
Key Innovation: Provided a comprehensive review and discussion of fluid flow governing equations, experimental/numerical methodologies, hydraulic property estimation (including REV and critical hydraulic gradient), and the significant effect of shear on fluid flow through single and complex fracture networks, with potential engineering applications.
53. Detecting Brick Kiln Infrastructure at Scale: Graph, Foundation, and Remote Sensing Models for Satellite Imagery Data
Core Problem: Large-scale monitoring of brick kiln infrastructure using satellite imagery is limited by sparse and outdated ground data.
Key Innovation: Proposing ClimateGraph, a region-adaptive graph-based model, and evaluating remote sensing and foundation models for scalable brick kiln detection from satellite imagery, highlighting complementary strengths.
54. High-Resolution Climate Projections Using Diffusion-Based Downscaling of a Lightweight Climate Emulator
Core Problem: Existing lightweight climate emulators provide accurate long-term statistics but at a coarse resolution (~300 km), which is inadequate for detailed regional impact assessments and subsequent geohazard analysis.
Key Innovation: Develops a deep learning-based downscaling framework using probabilistic diffusion-based generative models to downscale coarse climate emulator outputs to a 25 km resolution, preserving coarse-grained dynamics while generating fine-scaled climatological statistics.
55. A real-time UAS hyperspectral anomaly detection system
Core Problem: Obtaining real-time insights from hyperspectral anomaly detection on Uncrewed Aerial System (UAS) platforms is challenging due to post-processing requirements and UAS limitations.
Key Innovation: A novel, complete end-to-end real-time UAS hyperspectral anomaly detection system, integrating fast georectification and wireless transmission of concise anomaly information for immediate operator investigation, using relatively low-cost components.
56. MEMTS: Internalizing Domain Knowledge via Parameterized Memory for Retrieval-Free Domain Adaptation of Time Series Foundation Models
Core Problem: Time Series Foundation Models (TSFMs) degrade significantly in real-world domains due to temporal distribution shifts, and current adaptation solutions (DAPT, RAG) suffer from catastrophic forgetting or substantial retrieval overhead, creating scalability bottlenecks for real-time processing.
Key Innovation: MEMTS proposes a lightweight, plug-and-play method for retrieval-free domain adaptation in time series forecasting. Its Knowledge Persistence Module (KPM) internalizes domain-specific temporal dynamics into compact, learnable latent prototypes, enabling accurate, constant-time inference and near-zero latency adaptation without modifying the TSFM backbone.
57. Cast-R1: Learning Tool-Augmented Sequential Decision Policies for Time Series Forecasting
Core Problem: Traditional model-centric time series forecasting struggles in complex and evolving settings because most models lack the ability to autonomously acquire informative evidence, reason about future changes, or revise predictions through iterative decision processes.
Key Innovation: Cast-R1 reformulates time series forecasting as a sequential decision-making problem, introducing a memory-based state management mechanism and a tool-augmented agentic workflow. The agent autonomously interacts with a modular toolkit to extract features, invoke models, perform reasoning-based prediction, and iteratively refine forecasts, trained via supervised fine-tuning and multi-turn reinforcement learning.
58. Geometry-Aware Physics-Informed PointNets for Modeling Flows Across Porous Structures
Core Problem: Predicting flows through and around porous bodies is challenging due to coupled physics across fluid and porous regions and the need to generalize across diverse geometries and boundary conditions.
Key Innovation: Introduces Physics Informed PointNets (PIPN) and Physics Informed Geometry Aware Neural Operator (P-IGANO) to model these flows, enforcing Navier-Stokes and Darcy-Forchheimer equations within a unified loss, and conditioning networks on geometry and material parameters, demonstrating generalization to unseen shapes and variable conditions.
59. GeoEyes: On-Demand Visual Focusing for Evidence-Grounded Understanding of Ultra-High-Resolution Remote Sensing Imagery
Core Problem: Existing zoom-enabled Multimodal Large Language Models (MLLMs) for ultra-high-resolution (UHR) remote sensing Visual Question Answering (VQA) suffer from 'Tool Usage Homogenization,' limiting effective evidence acquisition for sparse and tiny task-relevant cues.
Key Innovation: GeoEyes, a staged training framework (UHR Chain-of-Zoom dataset and AdaZoom-GRPO reinforcement learning) that enables MLLMs to learn on-demand, evidence-grounded visual focusing with proper stopping behavior for UHR remote sensing imagery, significantly improving VQA accuracy.
60. Cross-view Domain Generalization via Geometric Consistency for LiDAR Semantic Segmentation
Core Problem: Existing domain generalization approaches for LiDAR semantic segmentation struggle in cross-view scenarios due to substantial differences in viewpoint-dependent structural incompleteness and non-uniform point density, limiting their real-world applicability.
Key Innovation: Introduction of CVGC (Cross-View Geometric Consistency), a novel framework that uses a cross-view geometric augmentation module to model viewpoint variations and a geometric consistency module to enforce consistent semantic and occupancy predictions across geometrically augmented point clouds, achieving state-of-the-art performance in cross-view domain generalization for LSS.
61. VariViT: A Vision Transformer for Variable Image Sizes
Core Problem: Vision Transformers (ViTs) are constrained to fixed-size input images, which leads to information degradation, artifacts, and suboptimal feature representation when dealing with variable-sized or irregularly shaped objects, particularly in domains like medical imaging.
Key Innovation: Proposes VariViT, an improved Vision Transformer designed to handle variable image sizes while maintaining a consistent patch size, utilizing a novel positional embedding resizing scheme and a new batching strategy, demonstrating enhanced feature representation and reduced computation time (up to 30%) compared to conventional architectures.
62. Pseudo-differential-enhanced physics-informed neural networks
Core Problem: Standard Physics-Informed Neural Networks (PINNs) can struggle with training efficiency, overall learning fidelity, and capturing high frequencies, especially in low collocation settings, due to issues like frequency bias.
Key Innovation: Introduces pseudo-differential enhanced PINNs, an extension of gradient enhancement in Fourier space, which improves training and learning fidelity, achieves superior PINN versus numerical error in fewer iterations, mitigates frequency bias by improving spectral eigenvalue decay of the neural tangent kernel (NTK), and accommodates fractional derivatives and advanced PINN techniques.
63. Parameter-Minimal Neural DE Solvers via Horner Polynomials
Core Problem: Traditional methods for solving differential equations with neural networks often require many parameters and may not enforce initial conditions exactly, limiting resource-efficient scientific modeling.
Key Innovation: Proposes a parameter-minimal neural architecture using Horner-factorized polynomials to solve differential equations, enforcing initial conditions exactly and employing a piecewise extension for accuracy, demonstrating high accuracy with very few parameters on ODE and heat-equation benchmarks.
64. Depth Completion as Parameter-Efficient Test-Time Adaptation
Core Problem: Adapting pre-trained 3D foundation models for depth completion using sparse geometric cues often involves training task-specific encoders that overfit and generalize poorly, leading to distortions in scene-specific measurements.
Key Innovation: Introduces CAPA, a parameter-efficient test-time optimization framework that adapts pre-trained 3D foundation models for depth completion by freezing the backbone and updating only minimal parameters (e.g., LoRA, VPT) using gradients from sparse observations, achieving state-of-the-art results and enforcing multi-frame consistency for videos.
65. Universal Algorithm-Implicit Learning
Core Problem: Current meta-learning methods are constrained to narrow task distributions with fixed feature and label spaces, limiting their practical applicability and hindering the development of truly universal meta-learners.
Key Innovation: Introduces a theoretical framework for universal meta-learning and presents TAIL, a transformer-based algorithm-implicit meta-learner that generalizes across varying domains, modalities, and label configurations, achieving state-of-the-art performance on few-shot benchmarks and demonstrating generalization to unseen domains and modalities with computational savings.
66. Wrivinder: Towards Spatial Intelligence for Geo-locating Ground Images onto Satellite Imagery
Core Problem: Accurately aligning ground-level imagery with geo-registered satellite maps is challenging due to large viewpoint gaps and unreliable GPS, hindering applications in mapping, navigation, and situational awareness.
Key Innovation: Introduced Wrivinder, a zero-shot, geometry-driven framework that aggregates multiple ground photographs to reconstruct a 3D scene and align it with overhead satellite imagery, achieving sub-30m geolocation accuracy. Also released MC-Sat, a benchmark dataset for this task.
67. PDE foundation models are skillful AI weather emulators for the Martian atmosphere
Core Problem: Developing skillful predictive weather emulators for complex atmospheric systems, such as the Martian atmosphere, is challenging due to limited training data or suitable compute budgets.
Key Innovation: Demonstrates that AI foundation models pretrained on diverse partial differential equations (PDEs) can be effectively adapted and fine-tuned to obtain skillful predictive weather emulators for the Martian atmosphere, extending a 2D model to 3D and performing well with sparse initial conditions.
68. Learning Gradient Flow: Using Equation Discovery to Accelerate Engineering Optimization
Core Problem: Traditional engineering optimization methods require expensive evaluations of the objective function and its gradient, hindering efficiency for complex problems.
Key Innovation: The Learned Gradient Flow (LGF) optimizer, which uses data-driven equation discovery to model and forecast continuous-time dynamics of unconstrained optimization problems, creating surrogate models to significantly expedite convergence by avoiding expensive evaluations.
69. Ambient Physics: Training Neural PDE Solvers with Partial Observations
Core Problem: Existing diffusion-based methods for reconstructing fields from partial observations require complete observations for training, which are often expensive, hazardous, or impossible to acquire in many scientific settings.
Key Innovation: Introduced Ambient Physics, a framework that enables learning the joint distribution of coefficient-solution pairs directly from partial observations for training neural PDE solvers, achieving state-of-the-art reconstruction performance with significantly fewer function evaluations by randomly masking already-observed measurements during training.
70. Distributed Quantum Gaussian Processes for Multi-Agent Systems
Core Problem: Classical Gaussian Processes are limited in expressivity and scalability for complex, large-scale real-world domains, and multi-agent systems require robust distributed optimization for quantum models.
Key Innovation: Proposes a Distributed Quantum Gaussian Process (DQGP) method for multi-agent systems, enhancing modeling capabilities and scalability by leveraging quantum computing and developing a Distributed consensus Riemannian Alternating Direction Method of Multipliers (DR-ADMM) algorithm, demonstrated on elevation datasets.
71. TKN: Transformer-based Keypoint Prediction Network For Real-time Video Prediction
Core Problem: Traditional video prediction methods prioritize accuracy over speed, making them unsuitable for real-time applications like danger prediction and warning due to complex models, redundant information, and sequential frame prediction.
Key Innovation: TKN, a transformer-based keypoint prediction network that extracts dynamic content unsupervised, uses an acceleration matrix for attention, and a parallel computing structure to achieve real-time video prediction at 1,176 fps, significantly reducing computation costs while maintaining performance.
72. A Survey on Generative Modeling with Limited Data, Few Shots, and Zero Shot
Core Problem: The challenge of applying generative modeling in real-world applications where large and diverse datasets are often unavailable, leading to issues like overfitting, frequency bias, and incompatible knowledge transfer.
Key Innovation: A comprehensive survey of Generative Modeling under Data Constraint (GM-DC), presenting a unified perspective on key challenges, introducing novel taxonomies for tasks and methodological approaches, reviewing over 230 papers, and highlighting future research directions relevant to data-constrained scenarios, including satellite imaging.
73. RainPro-8: An Efficient Deep Learning Model to Estimate Rainfall Probabilities Over 8 Hours
Core Problem: Radar-only deep learning models for precipitation forecasting have limitations such as short forecast lead times and inefficient integration of multiple data sources.
Key Innovation: Presents RainPro-8, an efficient deep learning model that integrates radar, satellite, and physics-based numerical weather prediction (NWP) data to provide high-resolution probabilistic precipitation forecasts over an 8-hour horizon in Europe, surpassing current operational systems.
74. Vision Transformers for Multi-Variable Climate Downscaling: Emulating Regional Climate Models with a Shared Encoder and Multi-Decoder Architecture
Core Problem: Global Climate Models (GCMs) have coarse spatial resolution, limiting their use in regional climate studies, while traditional Regional Climate Models (RCMs) are computationally expensive. Existing deep learning downscaling methods often focus on single variables, leading to inefficiencies and limited contextual awareness.
Key Innovation: Proposes a multi-variable Vision Transformer (ViT) with a shared encoder and variable-specific decoders (1EMD) to jointly downscale six key climate variables, achieving improved accuracy (5.5% MSE reduction) and computational efficiency (29-32% lower inference time per variable) compared to single-variable and other multi-variable baselines, effectively emulating RCM-scale downscaling.
75. Curriculum Multi-Task Self-Supervision Improves Lightweight Architectures for Onboard Satellite Hyperspectral Image Segmentation
Core Problem: High dimensionality of HSI data and slow satellite data transfer rates necessitate compact and efficient models for onboard processing to minimize redundant data transmission, while existing SSL methods may not efficiently capture both spatial and spectral reasoning for lightweight models.
Key Innovation: Introduces Curriculum Multi-Task Self-Supervised Learning (CMTSSL) for lightweight HSI architectures, integrating masked image modeling with decoupled spatial and spectral jigsaw puzzle solving, guided by curriculum learning to progressively increase data difficulty, enabling the encoder to jointly capture fine-grained spectral continuity, spatial structure, and global semantic features for efficient onboard satellite deployment.
76. Deep Learning for Clouds and Cloud Shadow Segmentation in Methane Satellite and Airborne Imaging Spectroscopy
Core Problem: Effective cloud and cloud shadow detection is a critical prerequisite for accurate retrieval of atmospheric methane or other trace gases in hyperspectral remote sensing, especially for high spatial resolution sensors like MethaneSAT and MethaneAIR, as these features bias methane retrievals and impact emission quantification.
Key Innovation: Deploys and evaluates conventional and advanced deep learning architectures (U-Net, SCAN) for cloud and cloud shadow detection in high spatial resolution methane satellite and airborne imaging spectroscopy data, demonstrating that deep learning models substantially improve detection quality, with U-Net excelling in spatial structure preservation and SCAN in fine boundary details, providing public data and code.
77. Geometry-to-Image Synthesis-Driven Generative Point Cloud Registration
Core Problem: Enhancing 3D point cloud registration performance by bridging 2D generative models with 3D matching tasks, specifically ensuring generated image pairs have both 2D-3D geometric consistency and cross-view texture consistency.
Key Innovation: Introduces Generative Point Cloud Registration, a paradigm that uses matching-specific controllable 2D generative models (DepthMatch-ControlNet and LiDARMatch-ControlNet) to synthesize cross-view consistent RGB images from depth maps or range maps, enabling geometry-color feature fusion for robust 3D registration.
78. Bias-Corrected Data Synthesis for Imbalanced Learning
Core Problem: Synthetic data generated for minority groups in imbalanced datasets can introduce bias, reducing prediction accuracy when treated as true data in classification problems.
Key Innovation: Proposes a bias correction procedure for synthetic data in imbalanced learning, providing consistent estimators for this bias by borrowing information from the majority group, thereby enhancing prediction accuracy and avoiding overfitting.
79. UK Hydrological Outlook using Historic Weather Analogues
Core Problem: Existing seasonal hydrological forecasting methods (ESP) have limitations in skill across different regions and seasons, particularly for winter river flow forecasts, hindering effective water resources planning and disaster risk reduction.
Key Innovation: The Historic Weather Analogues (HWA) method significantly improves winter river flow predictability nationally, especially for discriminating high and low flows, by leveraging climate information and exploring historically unseen weather sequences, offering a more skillful alternative to existing forecasting approaches in specific contexts.
80. Strain Localization and Failure Evolution in Rock–Coal Composites Considering Interaction Mechanisms: A Theoretical and Experimental Investigation
Core Problem: Understanding the synergistic bearing characteristics and failure mechanisms of roof-coal pillar composite structures is crucial to prevent coal pillar instability and dynamic disasters in deep mining.
Key Innovation: Developed a uniaxial compression mechanical model and conducted experiments to elucidate how rock-coal interaction modifies stress transfer, enhances local bearing capacity, and influences strain localization and failure evolution, providing a theoretical basis for preventing deep coal pillar instability.
81. Mechanical Characteristics of the Circular-Holed Granite Rock Under Static Coupled Dynamic Load: Insights for Dynamic Failure in a Deep Cavity
Core Problem: Understanding how static and dynamic loads interact to cause dynamic catastrophes and failure in deep cavities, particularly around stress concentrations, is essential for prevention.
Key Innovation: Experimentally investigated the dynamic mechanical properties and failure characteristics of circular-holed granite under static coupled dynamic loads using SHPB and DIC, revealing the role of static loads in strain energy accumulation and its impact on dynamic strength and failure modes, providing insights into rockburst mechanisms.
82. Strength Parameter Estimation from Triaxial Data Using the Úcar Failure Criterion
Core Problem: Accurate estimation of rock and concrete strength parameters (UCS, TS) from limited triaxial test data is challenging, particularly in early project stages or under budget constraints, hindering reliable geotechnical and rock engineering design.
Key Innovation: A methodology for fitting uniaxial compressive strength (UCS) and tensile strength (TS) using the nonlinear Úcar failure criterion from triaxial test data, enabling strength prediction for intact rock even with sparse datasets and providing general rock-type models.
83. STNet: A spatio-temporal network for binary change detection in high-resolution remote sensing images
Core Problem: The limitation of conventional single-temporal spatial modeling paradigms in change detection for remote sensing images, which fail to dynamically couple temporal changes with spatial context interdependencies.
Key Innovation: Proposes STNet, a novel spatio-temporal change detection network that jointly models temporal changes and spatial context. It includes Spatio-Temporal Relationship Modeling (STRM), Spatial Detail Preservation (SDP), and Neighbor Enhancement Decoder (NED), achieving state-of-the-art performance in multi-scale detection accuracy and generalization, particularly for irregular geometries and boundary continuity.
84. Benchmarking Atmospheric Circulation Variability in an AI Emulator, ACE2, and a Hybrid Model, NeuralGCM
Core Problem: Physics-based atmosphere-land models have biases in representing atmospheric variability, and new AI emulators and hybrid models require systematic evaluation against metrics grounded in fundamental atmospheric dynamics.
Key Innovation: Evaluated four atmospheric variability benchmarking metrics in a fully data-driven AI emulator (ACE2-ERA5) and a hybrid model (NeuralGCM). Found that both models capture large-scale tropical waves and extratropical eddy-mean flow interactions for short timescales, but struggle with slower timescales like the quasi-biennial oscillation (QBO) and Southern annular mode propagation.
85. Beyond Ground: Map-Free LiDAR Relocalization for UAVs
Core Problem: Existing LiDAR relocalization methods are primarily tailored for autonomous driving and exhibit degraded accuracy in UAV scenarios, especially with substantial yaw rotations and altitude variations, and current datasets lack real-world UAV flight characteristics.
Key Innovation: Proposes MAILS, a novel map-free LiDAR relocalization framework for UAVs, featuring a Locality-Preserving Sliding Window Attention module, coordinate-independent feature initialization, and a locally invariant positional encoding mechanism, along with a new large-scale UAV LiDAR localization dataset.
86. Sim2Radar: Toward Bridging the Radar Sim-to-Real Gap with VLM-Guided Scene Reconstruction
Core Problem: Learning-based millimeter-wave (mmWave) radar perception is bottlenecked by the scarcity and cost of collecting and annotating large-scale radar datasets, particularly for visually degraded indoor environments.
Key Innovation: Sim2Radar, an end-to-end framework that synthesizes training radar data directly from single-view RGB images by reconstructing material-aware 3D scenes using VLM-guided reasoning and simulating mmWave propagation, improving downstream 3D radar object detection via transfer learning.
87. Learning on the Fly: Replay-Based Continual Object Perception for Indoor Drones
Core Problem: Autonomous agents like indoor drones need to learn new object classes continually in real-time without catastrophic forgetting, but existing datasets are often outdoor-focused and Class-Incremental Learning (CIL) strategies need benchmarking for resource-constrained UAVs.
Key Innovation: Introduces a new indoor UAV video dataset with temporal coherence and benchmarks replay-based Class-Incremental Learning (CIL) strategies using YOLOv11-nano on resource-efficient drone platforms, demonstrating effective continual learning under tight memory budgets.
88. Federated Learning of Nonlinear Temporal Dynamics with Graph Attention-based Cross-Client Interpretability
Core Problem: Understanding temporal interdependencies across distributed industrial systems with heterogeneous, high-dimensional time series data is challenging in decentralized settings where raw measurements cannot be shared, client observations are heterogeneous, and local proprietary models cannot be modified.
Key Innovation: Presents a federated framework where clients map local observations to latent states, and a central server learns a graph-structured neural state transition model using a Graph Attention Network, providing the first interpretable characterization of cross-client temporal interdependencies via the Jacobian of the learned transition model and attention coefficients.
89. Joint Time Series Chain: Detecting Unusual Evolving Trend across Time Series
Core Problem: Existing time series chain (TSC) definitions only consider single time series, failing to detect unexpected evolving patterns in interrupted or across two related time series, which are crucial for identifying precursors of important events in complex systems.
Key Innovation: Introduces 'Joint Time Series Chain,' a new definition and effective ranking criterion specifically designed to find unusual evolving trends across interrupted or two related time series, mitigating robustness issues caused by gaps and outperforming existing TSC methods.
90. Optimized Certainty Equivalent Risk-Controlling Prediction Sets
Core Problem: Existing risk-controlling prediction sets (RCPS) provide probabilistic guarantees on expected risk but fail to capture tail behavior and worst-case scenarios crucial for safety-critical applications.
Key Innovation: Introduces Optimized Certainty Equivalent RCPS (OCE-RCPS), a novel framework providing high-probability guarantees on general optimized certainty equivalent (OCE) risk measures (like CVaR), leveraging upper confidence bounds to identify prediction set parameters that satisfy user-specified risk tolerance levels with provable reliability.
91. Joint Orientation and Weight Optimization for Robust Watertight Surface Reconstruction via Dirichlet-Regularized Winding Fields
Core Problem: Reconstructing robust watertight surfaces from unoriented point clouds with non-uniform sampling, noise, and outliers remains challenging, often requiring separate preprocessing steps.
Key Innovation: Dirichlet Winding Reconstruction (DiWR) is a robust method that jointly optimizes point orientations, per-point area weights, and confidence coefficients using the generalized winding number (GWN) field. It minimizes the Dirichlet energy of the induced winding field and incorporates GWN-based constraints to compensate for sampling issues, noise, and outliers, outperforming traditional multi-stage pipelines.
92. Testing For Distribution Shifts with Conditional Conformal Test Martingales
Core Problem: Existing conformal test martingales (CTMs) for detecting distribution shifts suffer from 'test-time contamination,' where post-shift observations dilute evidence, increasing detection delay and reducing power.
Key Innovation: Proposed a sequential test using conditional conformal test martingales (CTMs) that avoids contamination by comparing new samples to a fixed null reference dataset, providing anytime-valid type-I error control, asymptotic power one, and bounded expected detection delay, leading to faster shift detection.
93. KoopGen: Koopman Generator Networks for Representing and Predicting Dynamical Systems with Continuous Spectra
Core Problem: Representing and predicting high-dimensional, spatiotemporally chaotic dynamical systems, especially those with broadband or continuous spectra, remains challenging for data-driven models which often lack stability, interpretability, and scalability.
Key Innovation: Introduces KoopGen, a generator-based neural Koopman framework that models dynamics through a structured, state-dependent representation of Koopman generators, separating conservative transport from irreversible dissipation and enforcing operator-theoretic constraints.
94. Flow4R: Unifying 4D Reconstruction and Tracking with Scene Flow
Core Problem: Existing approaches for reconstructing and tracking dynamic 3D scenes often decouple geometry from motion, relying on separate models or explicit pose estimation.
Key Innovation: Proposes Flow4R, a unified framework that treats camera-space scene flow as the central representation linking 3D structure, object motion, and camera motion, predicting a minimal per-pixel property set from two-view inputs using a Vision Transformer.
95. EIDOS: Latent-Space Predictive Learning for Time Series Foundation Models
Core Problem: Most time series foundation models are pretrained by directly predicting future observations, leading to weakly structured latent representations that capture surface noise rather than coherent and predictable temporal dynamics.
Key Innovation: Introduces EIDOS, a foundation model family that shifts pretraining from future value prediction to latent-space predictive learning, training a causal Transformer to predict the evolution of latent representations, optimized via a joint objective integrating latent-space alignment, observational grounding, and direct forecasting supervision.
96. Restoration Adaptation for Semantic Segmentation on Low Quality Images
Core Problem: Semantic segmentation performance deteriorates on low-quality (LQ) images due to lack of clear semantic structures, and conventional image restoration models primarily focus on pixel-level fidelity, failing to recover task-relevant semantic cues for downstream tasks.
Key Innovation: Restoration Adaptation for Semantic Segmentation (RASS) is proposed, integrating semantic image restoration into the segmentation process. It includes a Semantic-Constrained Restoration (SCR) model that injects segmentation priors and transfers knowledge to segmentation via LoRA-based module merging and fine-tuning, enabling high-quality segmentation on LQ images.
97. Decentralized Federated Learning With Energy Harvesting Devices
Core Problem: Energy-intensive operations in decentralized federated learning (DFL) rapidly deplete limited device batteries, reducing operational lifetime and degrading learning performance, especially for edge devices.
Key Innovation: The paper applies energy harvesting to DFL systems, deriving a convergence bound and proposing a fully decentralized policy iteration algorithm for joint device scheduling and power control. This algorithm leverages local state information to accelerate convergence and improve sustainability for energy-constrained edge devices.
98. GeoFusionLRM: Geometry-Aware Self-Correction for Consistent 3D Reconstruction
Core Problem: Single-image 3D reconstruction with large reconstruction models (LRMs) often produces geometrically inconsistent and misaligned details, limiting fidelity.
Key Innovation: Introduces GeoFusionLRM, a geometry-aware self-correction framework that leverages the model's own normal and depth predictions, feeding back these geometric cues through a dedicated transformer and fusion module to refine structural accuracy and enforce consistency without additional supervision.
99. DenseMLLM: Standard Multimodal LLMs are Intrinsic Dense Predictors
Core Problem: Extending Multimodal Large Language Models (MLLMs) to fine-grained dense prediction tasks typically requires complex, task-specific decoders and customizations, increasing model complexity and limiting generalist design.
Key Innovation: Proposes DenseMLLM, a method that enables standard MLLMs to perform dense predictions (e.g., semantic segmentation, depth estimation) without additional task-specific decoders, using a novel vision token supervision strategy for multiple labels and tasks, demonstrating competitive performance across various benchmarks.
100. Learning Proposes, Geometry Disposes: A Modular Framework for Efficient Spatial Reasoning
Core Problem: It's an open question whether learning components should directly replace geometric estimation or serve as intermediate modules within pipelines for spatial perception (estimating camera motion and scene structure), and how to effectively integrate learning-based proposals with classical geometric algorithms for robust spatial reasoning.
Key Innovation: Proposes and investigates a modular framework for efficient spatial reasoning where learning proposes geometric hypotheses (e.g., pose and depth from VGGT) and geometric algorithms (e.g., classical point-to-plane RGB-D ICP) dispose estimation decisions, demonstrating that geometry is an essential arbiter for validating and absorbing learning-based geometric observations for robust spatial perception.
101. YOLO26: A Comprehensive Architecture Overview and Key Improvements
Core Problem: The continuous need to enhance object detection models (YOLO series) in terms of inference speed, performance on edge devices, and capabilities across various computer vision tasks like instance segmentation and pose estimation.
Key Innovation: Provides the first detailed architectural investigation of YOLO26, identifying key improvements such as the elimination of Distribution Focal Loss (DFL), implementation of End-to-End NMS-Free Inference, introduction of ProgLoss + Small-Target-Aware Label Assignment (STAL), and use of the MuSGD optimizer, which collectively boost inference speed (43% on CPU) and improve performance across multiple computer vision tasks.
102. Advances in Global Solvers for 3D Vision
Core Problem: Nonconvex geometric optimization problems in 3D vision are traditionally addressed by local or heuristic methods, lacking certifiable solutions and a unified understanding of global solvers.
Key Innovation: Provides the first systematic review of global solvers in geometric vision, unifying the field through a comprehensive taxonomy of Branch-and-Bound, Convex Relaxation, and Graduated Non-Convexity paradigms, analyzing their theoretical foundations, algorithmic designs, and practical enhancements across ten core vision tasks, and identifying future research directions.
103. It's a Matter of Time: Three Lessons on Long-Term Motion for Perception
Core Problem: The role and properties of long-term motion information for visual learning and perception tasks are not well understood, despite its potential importance.
Key Innovation: Leverages point-track estimation to demonstrate that long-term motion representations are rich in information (actions, objects, materials, spatial), generalize better than image representations in low-data/zero-shot settings, and offer a superior GFLOPs-accuracy trade-off.
104. SAILS: Segment Anything with Incrementally Learned Semantics for Task-Invariant and Training-Free Continual Learning
Core Problem: Existing continual learning methods for semantic segmentation suffer from repeated retraining, high computational costs, and catastrophic forgetting, limiting their real-world applicability.
Key Innovation: SAILS, a training-free framework for Class-Incremental Semantic Segmentation (CISS), decouples region extraction (using SAM) from semantic association, incorporates selective intra-class clustering, and avoids parameter updates to eliminate forgetting and achieve task-invariant performance, often surpassing training-based approaches.
105. GOT-JEPA: Generic Object Tracking with Model Adaptation and Occlusion Handling using Joint-Embedding Predictive Architecture
Core Problem: Recent generic object trackers lack robustness and generalization in unseen scenarios and have coarse occlusion reasoning, limiting their effectiveness in dynamic environments.
Key Innovation: GOT-JEPA, a model-predictive pretraining framework extending JEPA to predict tracking models, and OccuSolver, an occlusion perception module, enhance tracker generalization and robustness by providing stable pseudo supervision and detailed occlusion-pattern capture.
106. VIPA: Visual Informative Part Attention for Referring Image Segmentation
Core Problem: Existing Referring Image Segmentation methods struggle to effectively exploit visual contexts for fine-grained segmentation, leading to high-variance cross-modal projection and reduced semantic consistency.
Key Innovation: VIPA, a novel attention-based framework, leverages 'visual informative parts' (visual expressions) generated by a VEG module to provide structural and semantic visual target information, reducing cross-modal projection variance and enhancing semantic consistency for robust, fine-grained RIS.
107. Extending Multi-Source Bayesian Optimization With Causality Principles
Core Problem: Traditional Multi-Source Bayesian Optimization (MSBO) assumes independent input variables, limiting its effectiveness in scenarios where causal information is available and interventions can be performed.
Key Innovation: Proposes Multi-Source Causal Bayesian Optimization (MSCBO), a principled integration of MSBO and Causal Bayesian Optimization (CBO), leveraging causality principles to enhance optimization efficiency, reduce computational complexity, and improve convergence speed and scalability in higher-dimensional problems.
108. BEACONS: Bounded-Error, Algebraically-Composable Neural Solvers for Partial Differential Equations
Core Problem: Neural networks traditionally struggle to reliably generalize solutions for Partial Differential Equations (PDEs) beyond their training data, limiting their application in computational physics for extrapolatory regimes.
Key Innovation: Develops BEACONS, a framework for constructing formally-verified neural network PDE solvers with rigorous convergence, stability, and conservation properties, guaranteeing correctness even in extrapolatory regimes by using method of characteristics and compositional deep learning.
109. ThermEval: A Structured Benchmark for Evaluation of Vision-Language Models on Thermal Imagery
Core Problem: Vision-language models (VLMs) trained on RGB imagery fail to generalize to thermal images, which are critical for applications where visible light is insufficient, and existing benchmarks do not evaluate the necessary thermal perception and reasoning capabilities.
Key Innovation: Introduces ThermEval-B, a structured benchmark of approximately 55,000 thermal visual question answering pairs, including a new dataset (ThermEval-D) with dense per-pixel temperature maps, to specifically assess and drive progress in thermal vision language understanding.
110. Lang2Act: Fine-Grained Visual Reasoning through Self-Emergent Linguistic Toolchains
Core Problem: Existing Visual Retrieval-Augmented Generation (VRAG) frameworks for Vision-Language Models (VLMs) rely on rigid, pre-defined external tools and often decouple visual perception from subsequent reasoning, leading to potential loss of visual information.
Key Innovation: Lang2Act, a framework that enables fine-grained visual perception and reasoning through self-emergent linguistic toolchains, where VLMs self-explore and exploit high-quality actions as linguistic tools, significantly enhancing their visual perception capabilities through a two-stage Reinforcement Learning-based training.
111. Benchmarking Anomaly Detection Across Heterogeneous Cloud Telemetry Datasets
Core Problem: Most deep learning models for time-series anomaly detection are evaluated on single datasets, raising questions about their generalizability and ability to handle different types of telemetry in large-scale, high-dimensional cloud environments.
Key Innovation: This study benchmarks four deep learning models (GRU, TCN, Transformer, TSMixer) and Isolation Forest across four diverse cloud telemetry datasets using a unified training and evaluation pipeline. It demonstrates that anomaly detection performance is critically governed by calibration stability and feature-space geometry, beyond just model architecture.
112. Nonparametric Distribution Regression Re-calibration
Core Problem: Probabilistic regression models often produce overconfident predictions by prioritizing informativeness over calibration, and existing post-hoc re-calibration methods rely on weak notions of calibration or restrictive parametric assumptions.
Key Innovation: A novel nonparametric re-calibration algorithm based on conditional kernel mean embeddings, capable of correcting calibration error without restrictive modeling assumptions, and introducing an efficient characteristic kernel for real-valued targets.
113. Fast and accurate quasi-atom method for simultaneous atomistic and continuum simulation of solids
Core Problem: The computational expense and difficulty of simultaneously simulating solids at atomistic resolution in critical regions (e.g., cracks) and continuum scale in other parts, especially for complex multi-body potentials.
Key Innovation: A novel hybrid method for simultaneous atomistic and continuum simulation of solids using quasi-atoms for the continuum, with an online Machine Learning-based optimizer to match elastic properties, demonstrating significant computational speedup and accuracy for systems with Lennard-Jones and Tersoff potentials.
114. Activation-Space Uncertainty Quantification for Pretrained Networks
Core Problem: Reliable uncertainty estimates are crucial for deploying pretrained models, but many strong UQ methods require retraining, Monte Carlo sampling, or expensive computations, and may alter a frozen backbone's predictions.
Key Innovation: Introduces Gaussian Process Activations (GAPA), a post-hoc method that shifts Bayesian modeling to activations, preserving original point predictions while providing closed-form epistemic variances efficiently, matching or outperforming baselines in calibration and out-of-distribution detection across various tasks.
115. Permutation-based Inference for Variational Learning of Directed Acyclic Graphs
Core Problem: The fundamental challenge of estimating Bayesian network structures (DAGs) from observational data, particularly in causal discovery, due to difficulties in representing distributions over DAGs and estimating posteriors in combinatorial spaces.
Key Innovation: Introduction of PIVID, a method that jointly infers a distribution over permutations and DAGs using variational inference and continuous relaxations, demonstrating superior accuracy-uncertainty trade-offs and efficient scaling compared to deterministic and Bayesian approaches.
116. GraphFM: A generalist graph transformer that learns transferable representations across diverse domains
Core Problem: Graph neural networks (GNNs) typically require specialized models and significant hyperparameter tuning for each unique dataset, limiting their scalability and generalizability across diverse graph domains.
Key Innovation: Introduction of GraphFM, a scalable multi-graph pretraining approach using a Perceiver-based encoder with learned latent tokens to compress domain-specific features into a shared latent space, enabling the learning of transferable representations across 152 diverse graph datasets.
117. VCDF: A Validated Consensus-Driven Framework for Time Series Causal Discovery
Core Problem: Existing time series causal discovery methods are sensitive to noise, non-stationarity, and sampling variability, limiting their robustness and reliability in understanding dynamic systems.
Key Innovation: Proposes VCDF, a method-agnostic framework that improves robustness by evaluating the stability of causal relations across blocked temporal subsets, enhancing F1 scores and structural accuracy for time series causal discovery without altering base algorithms.
118. TRecViT: A Recurrent Video Transformer
Core Problem: Developing efficient and causal video modeling architectures that perform strongly on sparse and dense tasks while being computationally lightweight and suitable for real-time applications remains a challenge.
Key Innovation: Introduces TRecViT, a novel recurrent video transformer that uses a time-space-channel factorization with gated linear recurrent units for time, self-attention for space, and MLPs for channels, achieving state-of-the-art causal video modeling performance with significantly fewer parameters, lower memory footprint, and higher throughput.
119. Efficient Dual-domain Image Dehazing with Haze Prior Perception
Core Problem: Single-image dehazing methods, especially Transformer-based ones, are computationally expensive and often rely solely on spatial features, limiting their effectiveness in complex haze conditions and failing to adequately couple spatial and frequency domain information.
Key Innovation: Proposes DGFDNet, an efficient dual-domain dehazing network that explicitly aligns spatial and frequency degradation. It uses a Haze-Aware Frequency Modulator (HAFM) guided by dark channel priors for adaptive spectral filtering and a Multi-level Gating Aggregation Module (MGAM) for multi-scale feature fusion, achieving state-of-the-art performance, robustness, and real-time efficiency.
120. Zono-Conformal Prediction: Zonotope-Based Uncertainty Quantification for Regression and Classification Tasks
Core Problem: Current conformal prediction methods for uncertainty quantification are computationally expensive, data-intensive, and typically represent prediction sets with intervals, limiting their ability to capture dependencies in multi-dimensional outputs.
Key Innovation: Introduces zono-conformal prediction, a novel approach that constructs prediction zonotopes with assured coverage by placing zonotopic uncertainty sets directly into the base predictor model, identified via a single, data-efficient linear program, and less conservative than existing methods.
121. Challenges and Requirements for Benchmarking Time Series Foundation Models
Core Problem: Benchmarking Time Series Foundation Models (TSFMs) is challenging due to information leakage (train-test overlaps, temporal correlations) in existing evaluation studies, leading to inflated performance estimates.
Key Innovation: Identifies and categorizes information leakage issues in TSFM benchmarking and advocates for novel, principled evaluation methodologies to ensure reliable performance assessment.
122. Model-agnostic Selective Labeling with Provable Statistical Guarantees
Core Problem: Obtaining high-quality labels for large datasets is expensive, and AI models' predictions often contain unacceptable labeling errors, while existing selective labeling methods lack theoretical guarantees on AI-assigned label quality.
Key Innovation: Introduces "Conformal Labeling," a model-agnostic method that identifies trustworthy AI predictions by controlling the false discovery rate (FDR) using conformal p-values, providing theoretical guarantees for label quality.
123. MPCM-Net: Multi-scale network integrates partial attention convolution with Mamba for ground-based cloud image segmentation
Core Problem: Existing deep learning methods for ground-based cloud image segmentation have limitations in multi-scale context extraction, attention-based feature enhancement, and establishing global interdependencies in decoders, impacting accuracy and efficiency.
Key Innovation: Proposing MPCM-Net, a multi-scale network integrating Partial attention Convolutions with Mamba architectures, featuring an encoder with MPAC (MPC and MPA blocks) and a decoder with M2B and SSHD, to enhance segmentation accuracy and computational efficiency. Also introduces the CSRC dataset.
124. S2WMamba: A Spectral-Spatial Wavelet Mamba for Pansharpening
Core Problem: Fusing high-resolution panchromatic and low-resolution multispectral images to produce high-resolution multispectral images while disentangling spatial detail and spectral fidelity.
Key Innovation: Proposes S2WMamba, a model that explicitly disentangles frequency information using 2D/1D Haar DWT and performs lightweight cross-modal interaction with Mamba-based cross-modulation, achieving state-of-the-art pansharpening performance.
125. 3AM: 3egment Anything with Geometric Consistency in Videos
Core Problem: Video object segmentation methods struggle with geometric consistency under large viewpoint changes due to reliance on appearance features, while traditional 3D methods require expensive preprocessing and camera poses.
Key Innovation: 3AM, a training-time enhancement that integrates 3D-aware features from MUSt3R into SAM2 using a lightweight Feature Merger, achieving geometry-consistent recognition grounded in both spatial position and visual similarity from RGB input without camera poses or preprocessing.
126. LAViG-FLOW: Latent Autoregressive Video Generation for Fluid Flow Simulations
Core Problem: High-fidelity multiphase fluid flow simulations for applications like CO2 sequestration and geothermal production are computationally expensive, hindering uncertainty quantification and inversion.
Key Innovation: LAViG-FLOW, a latent autoregressive video generation diffusion framework, learns and extrapolates coupled saturation and pressure fields two orders of magnitude faster than traditional solvers, maintaining consistency and enabling efficient forecasting.
127. PhaSR: Generalized Image Shadow Removal with Physically Aligned Priors
Core Problem: Robust image shadow removal under diverse lighting conditions is challenging due to difficulty in disentangling illumination from reflectance and misaligned physical priors.
Key Innovation: PhaSR (Physically Aligned Shadow Removal) uses dual-level prior alignment, including Physically Aligned Normalization (PAN) for illumination correction and Geometric-Semantic Rectification Attention (GSRA) for harmonizing depth and semantic embeddings, achieving generalized and robust shadow removal.
128. Learning nonnegative matrix factorizations from compressed data
Core Problem: The computational cost and data access limitations associated with performing nonnegative matrix factorization (NMF) on large datasets, especially when direct access to the full original data is restricted.
Key Innovation: A flexible framework for scalable NMF that learns low-rank components directly from compressed (sketched) data, requiring only limited access to the original data, and demonstrating its effectiveness with adapted optimization problems and algorithms.
129. Denoising Diffusions with Optimal Transport: Localization, Curvature, and Multi-Scale Complexity
Core Problem: Understanding the fundamental mechanisms and limitations of denoising in diffusion-based generative models, particularly how the backward denoising process relates to optimal transport and what factors determine its difficulty and localization uncertainty.
Key Innovation: Proving that score denoising is the optimal backward map in transportation cost, identifying the curvature function as the determinant of localization uncertainty, and introducing a novel multi-scale curvature complexity measure that quantifies the difficulty of the denoising chain.
130. Multi-Spectral Gaussian Splatting with Neural Color Representation
Core Problem: Existing 3D Gaussian Splatting frameworks treat different spectral modalities separately, failing to exploit underlying spectral and spatial correlations for multi-spectral 3D reconstruction and rendering.
Key Innovation: Introduces MS-Splatting, a multi-spectral 3D Gaussian Splatting framework that uses a novel neural color representation to encode multi-spectral information into a compact, per-splat feature embedding, enabling joint learning of all bands and improving multi-spectral rendering quality.
131. AECBench: A Hierarchical Benchmark for Knowledge Evaluation of Large Language Models in the AEC Field
Core Problem: The robustness and reliability of Large Language Models (LLMs) in the specialized and safety-critical Architecture, Engineering, and Construction (AEC) domain remain inadequately evaluated.
Key Innovation: Establishing AECBench, a comprehensive, hierarchical benchmark with a five-level cognition-oriented evaluation framework, 23 tasks, a 4,800-question dataset, and an 'LLM-as-a-Judge' approach to quantify LLM strengths and limitations in AEC.
132. Conditionally Whitened Generative Models for Probabilistic Time Series Forecasting
Core Problem: Probabilistic forecasting of multivariate time series is challenging due to non-stationarity, inter-variable dependencies, and distribution shifts, with existing generative models often ignoring informative priors.
Key Innovation: Proposing Conditionally Whitened Generative Models (CW-Gen) that incorporate prior information through conditional whitening, using a novel Joint Mean-Covariance Estimator (JMCE) to improve predictive performance and mitigate distribution shift effects in time series forecasting.
133. BiasFreeBench: a Benchmark for Mitigating Bias in Large Language Model Responses
Core Problem: Existing studies on bias mitigation for Large Language Models (LLMs) use inconsistent baselines and metrics, and their evaluations often ignore the gap between probability-based assessments and real-world user interactions with LLM responses.
Key Innovation: Introducing BiasFreeBench, an empirical benchmark that unifies the evaluation of eight mainstream bias mitigation techniques across two test scenarios, and a response-level metric (Bias-Free Score) to measure the fairness, safety, and anti-stereotypicality of LLM outputs.
134. GuidedSampling: Steering LLMs Towards Diverse Candidate Solutions at Inference-Time
Core Problem: Repeated Sampling (RS) for Large Language Model (LLM) inference often struggles to generate diverse solution candidates, frequently relying on the same underlying approach and producing redundant samples.
Key Innovation: Proposing GuidedSampling, a new inference algorithm that decouples exploration and generation phases to increase the diversity of generated candidate solutions, leading to significant performance improvements and a more diverse set of concepts compared to traditional RS.
135. AgenticSciML: Collaborative Multi-Agent Systems for Emergent Discovery in Scientific Machine Learning
Core Problem: The design of Scientific Machine Learning (SciML) architectures, loss formulations, and training strategies remains an expert-driven, labor-intensive research process requiring extensive experimentation.
Key Innovation: Introduces AgenticSciML, a collaborative multi-agent system that autonomously proposes, critiques, and refines SciML solutions, discovering novel strategies that outperform human-designed baselines and single-agent systems across physics-informed and operator learning tasks.
136. Collaborative Spatiotemporal Anchors and Key Feature Enhancement for UAV Tracking
Core Problem: Target tracking in uncrewed aerial vehicle (UAV) platforms faces significant challenges due to high maneuverability-induced appearance variations (e.g., scale changes, viewpoint shifts) and error accumulation from fixed-interval sampling and dynamic reference updating.
Key Innovation: A novel dynamic spatiotemporal perception framework with error suppression, introducing spatiotemporal anchors that use Wasserstein-1 distance for geodesic-equidistant sampling of reference frames and a key feature enhancement module for attention-driven fusion of salient features, achieving state-of-the-art performance in UAV tracking robustness.
137. A Semi-analytical Criterion for a Toughness-Dominated Hydraulic Fracture Interacting with an Orthogonal Stratigraphic Interface
Core Problem: Uncontrolled vertical propagation of hydraulic fractures in layered reservoirs can compromise extraction efficiency and potentially induce aquifer contamination and fault reactivation.
Key Innovation: Developed a two-dimensional semi-analytical criterion for a toughness-dominated hydraulic fracture interacting with an orthogonal stratigraphic interface, predicting interaction behaviors (crossing, slippage, opening) based on elastic property contrast and stress conditions, aiding fracturing optimization.
138. Micromechanical Response of Laminated Shale Based on Nanoindentation and Microscratch Experiments with Numerical Simulation
Core Problem: A quantitative characterization of the micro-physical and mechanical properties of layered shale, including the influence of internal fabric and bedding on its mechanical response and crack propagation, was needed.
Key Innovation: Quantitative characterization of shale's elastic modulus, hardness, fracture toughness, and brittleness index using nanoindentation and microscratch experiments with numerical simulation, revealing anisotropic mechanical properties and crack propagation patterns relevant to shale oil exploitation.
139. Vulnerability assessment of storage tanks subject to fires under multi-source uncertainties using a novel ensemble-based non-parametric approach
Core Problem: Vulnerability assessment of storage tanks subject to fires is complicated by various multi-source uncertainties that can compromise the reliability and credibility of the results.
Key Innovation: A novel ensemble-based non-parametric approach that quantifies uncertainties in vulnerability assessment of storage tanks subject to fires using tolerance intervals, first obtaining one-sided tolerance intervals for time to failure and escalation probability, then two-sided tolerance intervals using an ensemble method, providing a robust uncertainty analysis tool.
140. Quantitative comparison of fine-tuning techniques for pretrained latent diffusion models in the generation of unseen SAR images
Core Problem: Adapting large pretrained latent diffusion models to generate very high-resolution Synthetic Aperture Radar (SAR) images, especially for rare or unseen scenes, while maintaining SAR imaging physics and prompt fidelity.
Key Innovation: Develops a framework for fine-tuning a text-to-image foundation model for SAR image generation, comparing full fine-tuning and LoRA techniques. Finds that a hybrid strategy best preserves SAR geometry and texture while maintaining prompt fidelity, enabling text-based control and multimodal conditioning for SAR scene data augmentation.
141. Drought-Induced damage detection in Iberian Scots Pine Forests through satellite remote sensing
Core Problem: The increasing intensity of drought events necessitates effective spatio-temporal monitoring systems to identify drought-induced forest decline and mortality, but understanding which satellite sensors and resolutions are optimal for this task is limited.
Key Innovation: Assessment of Landsat-8/OLI, Sentinel-2/MSI, and PlanetScope/SD sensors for detecting drought-induced forest decline, identifying Sentinel-2/MSI with SWIR or red edge bands as most performant (c. 80% correlation with canopy damage) and providing insights into optimal spatial resolution and forest cover effects.
142. Strength and modulus evolution in laboratory prepared deep soil mix samples and implications in design for high plasticity clays
Core Problem: Existing deep soil mixing (DSM) design approaches for high plasticity clays do not adequately account for the influence of initial soil consistency on strength evolution.
Key Innovation: Proposed a Consistency Dependent Mixing Efficacy (CDME) framework that links soil consistency (Liquidity Index), mixing efficiency, and achievable strength, demonstrating that maximum strength occurs within an intermediate consistency range and providing correlations for design.
143. Investigation into the dynamic compaction reinforcement characteristics and particle crushing effects of construction and demolition waste
Core Problem: Understanding the dynamic compaction effects and particle crushing in construction and demolition waste (CDW) is crucial for optimizing reinforcement and determining appropriate stopping criteria.
Key Innovation: Revealed irregular dynamic stress propagation and significant energy dissipation during compaction, proposed recommended stopping compaction counts for various energies, and developed a predictive model for particle crushing rate (Br) considering different compaction energies.
144. Environments Associated With Lightning Occurrence Based on Pre‐ and Post‐Convective Rawinsonde Measurements in Central Europe
Core Problem: A comprehensive understanding of the environmental factors and atmospheric parameters controlling lightning occurrence is needed for improved prediction.
Key Innovation: Evaluated 137,501 rawinsonde measurements and 327 convective parameters, identifying Lifted Index (LI), CAPE in a hail growth zone (CAPE_HGL), cold cloud depth, and equilibrium level temperature/height as the most robust predictors for lightning. The study also highlighted the role of buoyancy in sub-freezing temperatures and strong atmospheric flow in cold season lightning.
145. Three‐Dimensional Geostatistical Inverse Analyses of Transient Head and Temperature Data From a Long‐Term Heat Tracer Test
Core Problem: Improving the accuracy of subsurface heterogeneity characterization, specifically the distribution of hydraulic conductivity (K), which is crucial for understanding groundwater flow and contaminant transport.
Key Innovation: Investigated the performance of integrating transient head and temperature data in 3D geostatistical inverse analyses using the pilot point method. Results show that combining head and temperature data improves the prediction of heat tracer tests and provides non-redundant information for delineating subsurface K distribution.
146. Learning to Select Like Humans: Explainable Active Learning for Medical Imaging
Core Problem: Medical image analysis requires substantial labeled data, but expert annotation is expensive. Traditional active learning relies solely on predictive uncertainty, ignoring clinically meaningful features.
Key Innovation: An explainability-guided active learning framework that integrates spatial attention alignment into sample acquisition, combining classification uncertainty with attention misalignment to select informative samples, improving data efficiency and clinical interpretability.
147. Agentic Spatio-Temporal Grounding via Collaborative Reasoning
Core Problem: Spatio-Temporal Video Grounding (STVG) methods suffer from redundant computation, heavy supervision requirements, and limited generalization, especially in open-world and training-free scenarios.
Key Innovation: The Agentic Spatio-Temporal Grounder (ASTG) framework, which uses two specialized MLLM-based agents (Spatial Reasoning Agent and Temporal Reasoning Agent) to collaboratively and autonomously retrieve spatio-temporal tubes in videos, outperforming existing weakly-supervised and zero-shot approaches.
148. The Speed-up Factor: A Quantitative Multi-Iteration Active Learning Performance Metric
Core Problem: The evaluation of iterative active learning processes lacks appropriate quantitative performance metrics.
Key Innovation: Formally introducing the "speed-up factor," a quantitative multi-iteration active learning performance metric that indicates the fraction of samples needed to match random sampling performance, demonstrating its accuracy and stability.
149. Comparing Classifiers: A Case Study Using PyCM
Core Problem: Selecting an optimal classification model requires a robust and comprehensive understanding of its performance, but the choice of evaluation metrics can fundamentally shift the interpretation of a model's efficacy, potentially missing subtle performance trade-offs.
Key Innovation: Provides a tutorial on the PyCM library, demonstrating its utility for deep-dive evaluations of multi-class classifiers and emphasizing the necessity of a multi-dimensional evaluation framework to uncover subtle but important differences in model performance.
150. Frequency-Enhanced Hilbert Scanning Mamba for Short-Term Arctic Sea Ice Concentration Prediction
Core Problem: Vanilla Mamba models struggle with temporal correlations and boundary details in Arctic sea ice concentration (SIC) prediction.
Key Innovation: FH-Mamba, a Frequency-enhanced Hilbert scanning Mamba Framework, which introduces a 3D Hilbert scan for locality-preserving spatiotemporal traversal, incorporates wavelet transform for high-frequency details, and uses a Hybrid Shuffle Attention module to improve temporal consistency and edge reconstruction for Arctic SIC forecasting.
151. Fast Swap-Based Element Selection for Multiplication-Free Dimension Reduction
Core Problem: Standard dimension reduction methods like PCA rely on matrix multiplications, which can be a bottleneck on resource-constrained systems.
Key Innovation: A fast swap-based algorithm for multiplication-free element selection for dimension reduction, which efficiently determines optimal element subsets by deriving an objective change formula using the matrix inversion lemma, demonstrated on MNIST.
152. Out-of-Support Generalisation via Weight Space Sequence Modelling
Core Problem: Deep learning models frequently exhibit catastrophic failure and overconfident, unrealistic predictions when required to extrapolate on data points found outside their training distribution (out-of-support generalisation).
Key Innovation: Proposes WeightCaster, a framework that reformulates out-of-support generalisation as a sequence modelling task in the weight space, yielding plausible, interpretable, and uncertainty-aware predictions efficiently without explicit inductive biases.
153. Interpretable clustering via optimal multiway-split decision trees
Core Problem: Existing interpretable clustering methods based on binary decision trees suffer from high computational costs, suboptimal solutions, and excessively deep structures, making them difficult to interpret while maintaining accuracy.
Key Innovation: Proposes an interpretable clustering method based on optimal multiway-split decision trees, formulated as a 0-1 integer linear optimization problem, which is more tractable and integrates a one-dimensional K-means algorithm for flexible discretization, yielding concise decision rules and competitive performance.
154. Layer-Guided UAV Tracking: Enhancing Efficiency and Occlusion Robustness
Core Problem: Visual object tracking (VOT) for UAVs faces a significant challenge in balancing accuracy and efficiency, especially under unpredictable occlusion conditions.
Key Innovation: Introduces LGTrack, a unified UAV tracking framework that integrates dynamic layer selection, efficient feature enhancement (via Global-Grouped Coordinate Attention), and robust representation learning for occlusions (via Similarity-Guided Layer Adaptation), achieving state-of-the-art real-time speed and competitive tracking accuracy.
155. An Ensemble Learning Approach towards Waste Segmentation in Cluttered Environment
Core Problem: Accurate waste segmentation in complex, cluttered real-world environments (characterized by deformed, patternless, and overlapping objects) is challenging, hindering robotic waste localization and picking for recycling.
Key Innovation: An Ensemble Learning approach (EL-4) combining U-Net and FPN via a weighted average method significantly improves waste segmentation accuracy (IoU 0.8306) and reduces Dice loss, enhancing efficiency for robotic waste sorting at Material Recovery Facilities.
156. Physics Aware Neural Networks: Denoising for Magnetic Navigation
Core Problem: Airborne magnetic-anomaly navigation is challenged by aircraft-induced magnetic noise and the inadequacy of classical models (Tolles-Lawson) for stochastically corrupted magnetic data, requiring robust denoising for accurate geomagnetic field extraction.
Key Innovation: A framework based on physics-based constraints (divergence-free vector field, E(3)-equivariance) is proposed for denoising magnetic data using neural networks (specifically the Contiformer architecture), significantly improving predictive accuracy and physical plausibility for magnetic navigation.
157. Explore Intrinsic Geometry for Query-based Tiny and Oriented Object Detector with Momentum-based Bipartite Matching
Core Problem: Query-based object detectors struggle with tiny and oriented objects, particularly in aerial images, due to the underutilization of intrinsic geometry during feature decoding and inconsistencies caused by inter-stage bipartite matching.
Key Innovation: IGOFormer, a novel query-based oriented object detector, integrates an Intrinsic Geometry-aware Decoder to inject geometric embeddings for enhanced feature decoding and employs a Momentum-based Bipartite Matching scheme to stabilize inter-stage matching, achieving superior performance in aerial oriented object detection.
158. AnomaMind: Agentic Time Series Anomaly Detection with Tool-Augmented Reasoning
Core Problem: Existing time series anomaly detection methods struggle with context-dependent or diverse anomaly patterns due to a lack of adaptive feature preparation, reasoning-aware detection, and iterative refinement.
Key Innovation: Proposes AnomaMind, an agentic framework that reformulates anomaly detection as a sequential decision-making process, using a structured workflow with multi-turn tool interactions and self-reflection, supported by a hybrid inference mechanism with reinforcement learning for task-specific optimization.
159. Sufficient Conditions for Stability of Minimum-Norm Interpolating Deep ReLU Networks
Core Problem: Understanding the algorithmic stability and generalization error of deep neural networks, particularly minimum-norm interpolating deep ReLU networks, has been challenging within classical stability frameworks.
Key Innovation: Investigated sufficient conditions for the stability of minimum-norm interpolating deep ReLU networks, finding that stability is achieved when they contain a stable sub-network followed by a low-rank weight matrix layer, providing theoretical insights into generalization in overparameterized models.
160. A Multi-Agent Framework for Code-Guided, Modular, and Verifiable Automated Machine Learning
Core Problem: Traditional AutoML frameworks often function as 'black boxes' lacking flexibility and transparency, while LLM-based agents suffer from hallucinated logic and logic entanglement, leading to unreliable and unrecoverable failures in complex real-world engineering tasks.
Key Innovation: iML, a novel multi-agent framework for AutoML that employs Code-Guided Planning, Code-Modular Implementation, and Code-Verifiable Integration to eliminate hallucination, decouple components, and enforce physical feasibility, achieving superior performance and robustness on diverse real-world benchmarks.
161. Experiential Reinforcement Learning
Core Problem: Reinforcement learning for language models often struggles with sparse and delayed environmental feedback, making it challenging for LMs to implicitly infer how failures should translate into effective behavioral changes.
Key Innovation: Experiential Reinforcement Learning (ERL), a training paradigm that embeds an explicit experience-reflection-consolidation loop into the RL process, allowing models to generate attempts, receive feedback, reflect, and refine, thereby improving learning efficiency and final performance in sparse-reward environments and reasoning tasks.
162. QuRL: Efficient Reinforcement Learning with Quantized Rollout
Core Problem: The rollout process in Reinforcement Learning with Verifiable Rewards (RLVR) for Large Language Models (LLMs) is an efficiency bottleneck due to autoregressive decoding, consuming a significant portion of training time.
Key Innovation: Quantized Reinforcement Learning (QuRL), which uses a quantized actor to accelerate rollout during training, addressing challenges with Adaptive Clipping Range (ACR) to mitigate training collapse and invariant scaling to reduce quantization noise, achieving 20% to 80% faster rollout.
163. MarsRetrieval: Benchmarking Vision-Language Models for Planetary-Scale Geospatial Retrieval on Mars
Core Problem: Existing benchmarks for planetary science deep learning are limited to closed-set supervised visual tasks and lack support for text-guided retrieval for geospatial discovery on Mars.
Key Innovation: Introduces MarsRetrieval, a multimodal benchmark for evaluating vision-language models for Martian geospatial discovery, including paired image-text retrieval, landform retrieval, and global geo-localization tasks, and proposes a unified retrieval-centric protocol.
164. Explainability-Inspired Layer-Wise Pruning of Deep Neural Networks for Efficient Object Detection
Core Problem: The increasing complexity of deep neural networks for object detection poses challenges for deployment on resource-constrained platforms, and traditional magnitude-based pruning methods do not always align with the true functional contribution of network components.
Key Innovation: An explainability-inspired, layer-wise pruning framework is presented, leveraging SHAP-inspired gradient-activation attribution to estimate layer importance. This data-driven approach leads to improved accuracy-efficiency trade-offs for various object detection architectures compared to L1-norm-based methods.
165. Policy Gradient with Adaptive Entropy Annealing for Continual Fine-Tuning
Core Problem: Large pretrained vision models remain vulnerable to catastrophic forgetting when adapted to new tasks in class-incremental settings, even with parameter-efficient fine-tuning (PEFT), as most approaches rely on cross-entropy loss rather than directly minimizing misclassification error.
Key Innovation: Adaptive entropy annealing (aEPG) is proposed, a training strategy based on an Expected Policy Gradient (EPG) method that directly minimizes misclassification error. aEPG transitions from exploratory (CE-like) to exploitative (EPG-like) learning, outperforming CE-based methods and enhancing adaptation in pretrained vision models.
166. Multi-Agent Debate: A Unified Agentic Framework for Tabular Anomaly Detection
Core Problem: Tabular anomaly detection often relies on single detectors or static ensembles, which struggle with heterogeneous model families, distribution shifts, missingness, and rare anomalies, leading to performance issues and lack of robustness.
Key Innovation: MAD, a Multi-Agent Debating framework that treats disagreement among heterogeneous ML detectors as a first-class signal, resolving it through a mathematically grounded coordination layer with LLM-based critics. It provides improved robustness, auditable traces, and regret guarantees.
167. Radial-VCReg: More Informative Representation Learning Through Radial Gaussianization
Core Problem: Existing self-supervised learning methods like VCReg, which regularize only first and second-order feature statistics, cannot fully achieve maximum entropy, hindering the learning of maximally informative representations due to the curse of dimensionality.
Key Innovation: Radial-VCReg, which augments VCReg with a radial Gaussianization loss that aligns feature norms with the Chi distribution, transforming a broader class of distributions towards normality. This improves performance by reducing higher-order dependencies and promoting more diverse and informative representations.
168. Integrating Unstructured Text into Causal Inference: Empirical Evidence from Real Data
Core Problem: Causal inference, a critical tool for informing business decisions, traditionally relies heavily on structured data, which is often incomplete or unavailable in many real-world scenarios, limiting its applicability.
Key Innovation: A framework that leverages transformer-based language models to perform causal inference using unstructured text, demonstrating consistent results with those obtained from structured data. This extends causal inference to scenarios where only textual data is available.
169. Whom to Query for What: Adaptive Group Elicitation via Multi-Turn LLM Interactions
Core Problem: Eliciting information to reduce uncertainty about latent group-level properties from surveys is challenging due to limited questioning effort, real costs, and missing data, and existing methods don't adapt respondent selection or leverage population structure effectively.
Key Innovation: A theoretically grounded framework for adaptive group elicitation that combines an LLM-based expected information gain objective for scoring questions with heterogeneous graph neural network propagation for imputing missing responses and guiding per-round respondent selection. This improves population-level response prediction under constrained budgets.
170. Machine Learning as a Tool (MLAT): A Framework for Integrating Statistical ML Models as Callable Tools within LLM Agent Workflows
Core Problem: Conventional LLM pipelines treat ML inference as a static preprocessing step, limiting the LLM's ability to dynamically decide when and how to use quantitative predictions based on conversational context.
Key Innovation: Machine Learning as a Tool (MLAT), a design pattern that exposes pre-trained statistical ML models as callable tools within LLM agent workflows, allowing the orchestrating agent to invoke quantitative predictions and reason about their outputs in context. This is demonstrated with PitchCraft, a system for generating sales proposals with ML-predicted pricing.
171. Differential pose optimization in descriptor space -- Combining Geometric and Photometric Methods for Motion Estimation
Core Problem: The fundamental problem of two-frame relative pose optimization involves a trade-off between accuracy, robustness, and loop closing when using either photometric or re-projection errors, which are tied to specific feature paradigms.
Key Innovation: Proposes a unified approach combining geometric and photometric methods by replacing photometric error with a descriptor residual from densely sampled geometric feature descriptors, aiming for sub-pixel accuracy and expressiveness in motion estimation.
172. A Study on Multi-Class Online Fuzzy Classifiers for Dynamic Environments
Core Problem: Conventional online fuzzy classifiers are limited to two-class problems, and there is a need to extend them for multi-class classification in dynamic environments where data arrives sequentially.
Key Innovation: Proposes and evaluates an extension of conventional online fuzzy classifiers to handle multi-class problems in dynamic environments, where fuzzy if-then rules are learned sequentially from incoming data, demonstrating its performance on synthetic and benchmark datasets.
173. Gaussian Mesh Renderer for Lightweight Differentiable Rendering
Core Problem: Traditional mesh-based differentiable renderers suffer from slow or heavy optimization, despite triangle mesh models being popular for surface reconstruction, while 3D Gaussian Splatting (3DGS) offers fast rendering but is not directly a mesh representation.
Key Innovation: Proposes Gaussian Mesh Renderer (GMR), a lightweight differentiable mesh renderer that integrates Gaussian and mesh representations, analytically deriving Gaussian primitives from mesh triangles to preserve structural fidelity and enable smoother gradients for better optimization.
174. Deep Image Prior for Computed Tomography Reconstruction
Core Problem: Conventional deep learning methods for image reconstruction, particularly in computed tomography, rely on large, supervised datasets which are often unavailable, and struggle with noise and efficiency.
Key Innovation: Provides a comprehensive overview of the Deep Image Prior (DIP) framework for CT reconstruction, demonstrating its unsupervised capability with single measurements, outlining key algorithmic choices, and reviewing strategies for overfitting mitigation and computational improvements, validated on real 5CT data.
175. Orthogonalized Multimodal Contrastive Learning with Asymmetric Masking for Structured Representations
Core Problem: Existing multimodal contrastive learning methods primarily capture redundant cross-modal signals, often neglecting modality-specific and interaction-driven information, leading to incomplete or entangled representations.
Key Innovation: Introduces COrAL, a principled framework with a dual-path architecture and orthogonality constraints to explicitly disentangle and simultaneously preserve redundant, unique, and synergistic information within multimodal representations, using asymmetric masking to promote synergy modeling.
176. Image Generation with a Sphere Encoder
Core Problem: The high computational cost and multi-step nature of existing state-of-the-art image generation models, such as diffusion models.
Key Innovation: Introducing the Sphere Encoder, an efficient generative framework capable of producing images in a single forward pass by learning an encoder that maps natural images uniformly onto a spherical latent space and a decoder that maps random latent vectors back to image space.
177. EditCtrl: Disentangled Local and Global Control for Real-Time Generative Video Editing
Core Problem: The computational inefficiency of existing high-fidelity generative video editing methods, which often process full video context regardless of the inpainting mask's size, even for sparse, localized edits.
Key Innovation: EditCtrl, an efficient video inpainting control framework that focuses computation only where needed, featuring a novel local video context module operating solely on masked tokens and a lightweight temporal global context embedder for video-wide consistency, leading to significant computational efficiency and improved editing quality.
178. TemporalBench: A Benchmark for Evaluating LLM-Based Agents on Contextual and Event-Informed Time Series Tasks
Core Problem: It is unclear whether strong forecasting performance in LLM-based agents reflects genuine temporal understanding or merely the ability to reason under contextual and event-driven conditions, as existing benchmarks do not adequately diagnose these capabilities.
Key Innovation: TemporalBench is introduced as a multi-domain benchmark with a four-tier task taxonomy to evaluate temporal reasoning behavior (historical structure, context-free, contextual, and event-conditioned prediction) in LLM-based agents. It reveals that strong numerical forecasting accuracy does not reliably translate into robust contextual or event-aware temporal reasoning.
179. Graph neural networks uncover structure and functions underlying the activity of simulated neural assemblies
Core Problem: Existing machine learning approaches for analyzing complex heterogeneous systems, such as recurrent neural networks and transformers, emphasize predictive accuracy but offer limited interpretability regarding the underlying mechanisms governing their activity.
Key Innovation: A graph neural network (GNN) framework, trained to predict observable dynamics, is applied to simulated neural assemblies. This method jointly reveals the connectivity matrix, neuron types, signaling functions, and in some cases, hidden external stimuli, providing both reliable forecasts of neural activity and interpretable decomposition of governing mechanisms.
180. Stochastic variance reduced extragradient methods for solving hierarchical variational inequalities
Core Problem: Solving optimization problems formulated as hierarchical variational inequalities, especially with finite-sum representation of smooth operators, lacks established convergence rates and complexity statements for variance-reduced stochastic algorithms.
Key Innovation: The first work to prove convergence rates and complexity statements for variance-reduced stochastic algorithms approaching the solution of hierarchical variational inequalities in both Euclidean and Bregman setups.
181. Differentiable Rule Induction from Raw Sequence Inputs
Core Problem: Most differentiable Inductive Logic Programming (ILP) methods rely on symbolic datasets and struggle to learn rules directly from raw, continuous data due to the 'explicit label leakage' problem when mapping inputs to symbolic variables.
Key Innovation: Integrates a self-supervised differentiable clustering model with a novel differentiable ILP model, enabling rule learning from raw time series and image data without explicit label leakage, leading to intuitive and precise generalized rules.
182. A Unified Physics-Informed Neural Network for Modeling Coupled Electro- and Elastodynamic Wave Propagation Using Three-Stage Loss Optimization
Core Problem: Solving coupled time-dependent partial differential equations (PDEs) for complex physical systems like electro-elastodynamics is challenging with traditional numerical methods.
Key Innovation: Developed a Physics-Informed Neural Network (PINN) model with a three-stage loss optimization for a one-dimensional coupled electro-elastodynamic system, demonstrating its effectiveness as a mesh-free solver for such coupled time-dependent PDE systems.
183. High-fidelity 3D reconstruction for planetary exploration
Core Problem: Traditional 3D reconstruction techniques (SfM, SLAM) for planetary exploration struggle to capture radiometric detail or scale efficiently in unstructured, low-texture extraterrestrial environments, hindering onboard spatial awareness for autonomous systems.
Key Innovation: Integrates radiance field-based methods (NeRF, Gaussian Splatting) into a unified, automated environment reconstruction pipeline for planetary robotics, generating dense, photorealistic, and metrically consistent 3D representations from minimal visual input.
184. A Hybrid TGN-SEAL Model for Dynamic Graph Link Prediction
Core Problem: Predicting links in sparse, continuously evolving networks is challenging, as conventional methods and Temporal Graph Networks (TGNs) struggle to capture both structural and temporal dependencies accurately, especially under sparse conditions and with transient interactions.
Key Innovation: Improves the TGN framework by extracting enclosing subgraphs around candidate links, integrating local topology to jointly learn structural and temporal information. This hybrid TGN-SEAL approach increases average precision by 2.6% over standard TGNs on sparse datasets, demonstrating more robust link prediction in dynamic networks.
185. Federated Ensemble Learning with Progressive Model Personalization
Core Problem: Personalized Federated Learning (PFL) faces a fundamental tradeoff: deep shared components hinder personalization, while large local heads lead to overfitting under limited client data, especially with rigid, shallow personalization components.
Key Innovation: Proposes a boosting-inspired framework that constructs an ensemble of models for each client, progressively increasing the depth of personalized components while controlling complexity via low-rank factorization or width shrinkage. This design limits overfitting, reduces per-client bias, and provides theoretical generalization bounds, outperforming state-of-the-art PFL methods on various datasets.
186. Fast Compute for ML Optimization
Core Problem: Machine learning optimization often requires user-specified learning-rate and momentum schedules, and can be slow for ill-conditioned problems, especially when exploring regularization paths.
Key Innovation: Introduces the Scale Mixture EM (SM-EM) algorithm, derived from a variance-mean scale-mixture representation of losses, which removes the need for user-specified schedules. With Nesterov acceleration, SM-EM achieves up to 13x lower final loss than tuned Adam on ill-conditioned logistic regression and a 10x runtime reduction for regularization paths by sharing sufficient statistics.
187. CAIRO: Decoupling Order from Scale in Regression
Core Problem: Standard regression methods conflate the learning of ordering with the learning of scale, rendering models vulnerable to outliers and heavy-tailed noise.
Key Innovation: CAIRO (Calibrate After Initial Rank Ordering), a framework that decouples regression into two stages: learning a scoring function via a scale-invariant ranking loss, followed by target scale recovery via isotonic regression, improving robustness in noisy regimes.
188. Frequentist Regret Analysis of Gaussian Process Thompson Sampling via Fractional Posteriors
Core Problem: Existing frequentist regret analyses of Gaussian Process Thompson Sampling (GP-TS) often rely on domain discretization, limiting their generality and applicability across different kernel classes.
Key Innovation: A unified, discretization-free frequentist regret analysis framework for GP-TS based on fractional Gaussian process posteriors, providing kernel-agnostic bounds that apply broadly across various kernel classes.
189. Constrained and Composite Sampling via Proximal Sampler
Core Problem: Efficiently sampling from log-concave distributions under convex constraints or with composite structures, where enforcing feasibility or handling composite functions without degrading mixing is challenging.
Key Innovation: Development of a proximal sampler-based approach for constrained and composite sampling problems, leveraging epigraph transformations and minimal oracle access to enforce feasibility and handle composite structures, with established mixing time bounds.
190. GenPANIS: A Latent-Variable Generative Framework for Forward and Inverse PDE Problems in Multiphase Media
Core Problem: Inverse problems and inverse design in multiphase media are challenging due to discrete-valued material fields, making them non-differentiable and incompatible with gradient-based methods, often requiring separate heavyweight models for forward and inverse tasks.
Key Innovation: GenPANIS, a unified latent-variable generative framework that preserves exact discrete microstructures while enabling gradient-based inference through continuous latent embeddings, supporting bidirectional inference and training with unlabeled data/physics residuals, and outperforming state-of-the-art methods with fewer parameters and uncertainty quantification.
191. Learning with Subset Stacking
Core Problem: Developing a regression algorithm that effectively handles heterogeneous relationships between input and output variables across the predictor space.
Key Innovation: Introduction of LESS (LEarning with Subset Stacking), a new regression algorithm that generates subsets, trains local predictors for each, and combines them in a novel way, showing competitive performance against state-of-the-art methods.
192. Cross-Modal Mapping: Mitigating the Modality Gap for Few-Shot Image Classification
Core Problem: The 'modality gap' in pre-trained visual-language models (e.g., CLIP) leads to suboptimal performance in few-shot image classification because image and text features have inconsistent distributions in the joint embedding space, hindering their direct use as class prototypes.
Key Innovation: Proposes Cross-Modal Mapping (CMM), a method that globally aligns image features with the text feature space via linear transformation and optimizes local spatial relationships using triplet loss, significantly enhancing cross-modal consistency and improving few-shot image classification accuracy across various datasets.
193. SWIFT: Mapping Sub-series with Wavelet Decomposition Improves Time Series Forecasting
Core Problem: Existing lightweight models for Long-term Time Series Forecasting (LTSF) perform poorly on non-stationary sequences, and large models are computationally expensive for resource-constrained environments.
Key Innovation: Introduces SWIFT, a lightweight and efficient LTSF model that uses wavelet transform for lossless downsampling, a learnable filter for cross-band information fusion, and a shared linear layer/shallow MLP for sub-series mapping, achieving state-of-the-art performance with significantly fewer parameters.
194. Simulating the Real World: A Unified Survey of Multimodal Generative Models
Core Problem: Existing multimodal generative models for real-world simulation often treat different modalities (2D, video, 3D, 4D) independently and lack a unified framework for systematically integrating their connections.
Key Innovation: Presents the first unified survey of multimodal generative models, systematically integrating 2D, video, 3D, and 4D generation within a single framework to advance the study of real-world simulation, providing a comprehensive review of datasets, metrics, and future directions.
195. Measure Twice, Cut Once: A Semantic-Oriented Approach to Video Temporal Localization with Video LLMs
Core Problem: Recent video LLM methods for temporal localization struggle to leverage LLMs' pre-trained semantic understanding capabilities due to their focus on uninformative timestamp outputs.
Key Innovation: Proposes MeCo, a timestamp-free, semantic-oriented framework that fine-tunes video LLMs using structural token generation, query-focused captioning, and a contrastive learning-driven grounding module to achieve holistic temporal segmentation and outperform timestamp-based methods for video temporal localization.
196. Heuristic Methods are Good Teachers to Distill MLPs for Graph Link Prediction
Core Problem: Existing GNN-to-MLP distillation methods for graph link prediction overlook the potential of specialized models and weaker heuristic methods as teachers, leading to suboptimal performance and efficiency.
Key Innovation: Discovers that heuristic methods can be surprisingly effective teachers for distilling MLPs for graph link prediction, and proposes Ensemble Heuristic-Distilled MLPs (EHDM) which integrates complementary signals, achieving significant performance improvement (7.93%) and reduced training time (1.95-3.32x) over previous GNN-to-MLP approaches.
197. S2R-HDR: A Large-Scale Rendered Dataset for HDR Fusion
Core Problem: The generalization of learning-based high dynamic range (HDR) fusion models is often limited by the scarcity of large-scale, real-world HDR training data from dynamic scenes, which is costly and challenging to collect.
Key Innovation: Introduces S2R-HDR, the first large-scale, high-quality synthetic dataset (24,000 HDR samples) for HDR fusion, generated using Unreal Engine 5, and proposes S2R-Adapter for domain adaptation to bridge the synthetic-to-real gap, achieving state-of-the-art performance on real-world datasets.
198. 3DRot: Rediscovering the Missing Primitive for RGB-Based 3D Augmentation
Core Problem: RGB-based 3D tasks suffer from scarce annotations and a limited augmentation toolbox because many image transforms (like rotations) disrupt geometric consistency, leading to a lack of rigorous 3D rotation augmentation.
Key Innovation: Introduces 3DRot, a plug-and-play augmentation that rotates and mirrors images about the camera's optical center while synchronously updating RGB images, camera intrinsics, object poses, and 3D annotations to preserve projective geometry without relying on scene depth, improving performance in monocular 3D detection and depth estimation.
199. ART: Adaptive Resampling-based Training for Imbalanced Classification
Core Problem: Traditional resampling methods for imbalanced classification use fixed distributions, failing to adapt to changes in class-wise learning difficulty, which limits overall model performance.
Key Innovation: Proposes Adaptive Resampling-based Training (ART), which periodically updates the training data distribution based on class-wise macro F1 scores, allowing the model to incrementally shift attention to underperforming classes and consistently outperforming existing resampling and algorithm-level methods on diverse imbalanced classification tasks.
200. Multi-scale Autoregressive Models are Laplacian, Discrete, and Latent Diffusion Models in Disguise
Core Problem: Understanding the design choices that drive the quality-efficiency trade-off in Visual Autoregressive (VAR) models and establishing their relationship to other generative models like diffusion models.
Key Innovation: Reinterpreting VAR models as iterative refinement models with a deterministic forward process building a Laplacian-style latent pyramid and a learned backward process, explicitly linking them to denoising diffusion, and identifying key modeling choices (latent space refinement, discrete prediction, spatial frequency decomposition) that contribute to their efficiency and quality, with potential applications in weather forecasting.
201. PAGCNet: A Pose-Aware and Geometry Constrained Framework for Panoramic Depth Estimation
Core Problem: Reconstructing room background depth as a strong geometric constraint for panoramic depth estimation in complex indoor scenes without external measurements remains an open challenge.
Key Innovation: PAGCNet, a pose-aware and geometry-constrained framework, is proposed. It jointly estimates room layout, camera pose, depth, and region segmentation, using a PA-BDR component to resolve camera pose and compute background depth as a geometric prior, which is then adaptively fused with initial depth predictions.
202. AnyUp: Universal Feature Upsampling
Core Problem: Existing learning-based feature upsamplers require re-training for each feature extractor, limiting their generalization to different feature types and resolutions.
Key Innovation: Proposes AnyUp, an inference-time feature-agnostic upsampling architecture that achieves state-of-the-art results, generalizes across feature types, and preserves semantics efficiently.
203. Top-Down Semantic Refinement for Image Captioning
Core Problem: Large Vision-Language Models (VLMs) struggle with image captioning to maintain global narrative coherence while capturing rich details due to their myopic single-step generation process.
Key Innovation: Proposes Top-Down Semantic Refinement (TDSR), a framework that redefines image captioning as a goal-oriented hierarchical refinement planning problem using an MDP and an efficient MCTS algorithm, significantly enhancing VLM performance in fine-grained description and compositional generalization.
204. Fourier-RWKV: A Multi-State Perception Network for Efficient Image Dehazing
Core Problem: Image dehazing under real-world non-uniform haze conditions, where Transformer-based methods have high computational complexity hindering real-time deployment.
Key Innovation: Proposes Fourier Receptance Weighted Key Value (Fourier-RWKV), a novel dehazing framework with linear complexity, integrating Spatial-form Perception, Frequency-domain Perception, and Semantic-relation Perception to achieve state-of-the-art performance efficiently.
205. NeRV360: Neural Representation for 360-Degree Videos with a Viewport Decoder
Core Problem: Applying implicit neural representations (NeRV) to high-resolution 360-degree videos results in high memory usage and slow decoding, making real-time applications impractical.
Key Innovation: NeRV360, an end-to-end framework that decodes only the user-selected viewport instead of the entire panoramic frame, integrating viewport extraction into decoding and using a spatial-temporal affine transform module for conditional decoding.
206. CliffordNet: All You Need is Geometric Algebra
Core Problem: Modern computer vision architectures rely on stacking heuristic modules (spatial/channel mixers), lacking a unified interaction mechanism grounded in mathematical first principles.
Key Innovation: Clifford Algebra Network (CAN) or CliffordNet, a vision backbone grounded purely in Geometric Algebra, using the Clifford Geometric Product as a unified interaction mechanism to capture feature coherence and structural variation, leading to efficient and accurate models without Feed-Forward Networks.
207. Q-Hawkeye: Reliable Visual Policy Optimization for Image Quality Assessment
Core Problem: Existing RL-based Image Quality Assessment (IQA) methods suffer from reliability limitations, including uniform advantage weighting that amplifies noisy signals from unstable predictions and an overemphasis on text-grounded reasoning over visual perception.
Key Innovation: Q-Hawkeye, an RL-based framework, introduces Uncertainty-Aware Dynamic Optimization (reweighting samples based on predictive uncertainty) and Perception-Aware Optimization (using an Implicit Perception Loss with paired degraded/original images) to enhance the reliability and generalization of IQA.
208. A Meta-Knowledge-Augmented LLM Framework for Hyperparameter Optimization in Time-Series Forecasting
Core Problem: Hyperparameter optimization (HPO) for deep learning models, particularly in time-series forecasting, is computationally expensive, difficult to interpret, and typically treats tuning tasks independently, offering limited insight into decisions.
Key Innovation: Introduces LLM-AutoOpt, a hybrid HPO framework combining Bayesian Optimization with LLM-based contextual reasoning, which encodes structured meta-knowledge (dataset meta-features, model descriptions, historical outcomes) to enable context-aware, stable, and interpretable hyperparameter refinement for time-series forecasting.
209. Exploring Real-Time Super-Resolution: Benchmarking and Fine-Tuning for Streaming Content
Core Problem: Existing real-time super-resolution methods struggle with the unique characteristics of compressed video content from streaming media, and current datasets do not accurately reflect real-world streaming scenarios, limiting benchmark relevance.
Key Innovation: Introduced StreamSR, a comprehensive dataset sourced from YouTube for streaming content, and benchmarked 11 state-of-the-art real-time super-resolution models. Proposed EfRLFN, an efficient real-time model integrating Efficient Channel Attention and a hyperbolic tangent activation function, demonstrating improved visual quality and runtime performance, and showing significant gains from fine-tuning on StreamSR.
210. Zooming without Zooming: Region-to-Image Distillation for Fine-Grained Multimodal Perception
Core Problem: Multimodal Large Language Models (MLLMs) struggle with fine-grained perception where decisive evidence is small, and existing 'Thinking-with-Images' methods incur high latency due to iterative zooming during inference.
Key Innovation: Proposed Region-to-Image Distillation, which transforms inference-time zooming into a training-time primitive by distilling region-grounded supervision from strong teacher models to a student MLLM. Also introduced ZoomBench, a hybrid-annotated benchmark, demonstrating leading performance on fine-grained perception benchmarks without tool use.
211. Reliable Thinking with Images
Core Problem: Existing Thinking with Images (TWI) methods for Multi-modal Large Language Models (MLLMs) assume faultless interleaved image-text Chain-of-Thought (CoT), which is often violated by 'Noisy Thinking' (imperfect visual cues and answer reasoning), leading to error accumulation and degraded performance.
Key Innovation: Proposed Reliable Thinking with Images (RTWI) to address the Noisy Thinking problem. RTWI estimates the reliability of visual cues and textual CoT in a unified text-centric manner and employs robust filtering and voting modules to prevent error accumulation, demonstrating effectiveness across seven benchmarks.
212. Deep Two-Way Matrix Reordering for Relational Data Analysis
Core Problem: Most existing matrix reordering techniques for relational data analysis rely on predefined feature representations or prior knowledge about structural patterns, which is not always available for observed matrices.
Key Innovation: Proposed Deep Two-Way Matrix Reordering (DeepTMR), a neural network-based method that automatically extracts nonlinear row/column features from an observed matrix for reordering. It also provides a denoised mean matrix as an output to visualize the global structure of the reordered matrix, demonstrating effectiveness on synthetic and practical datasets.
213. When Attention Collapses: How Degenerate Layers in LLMs Enable Smaller, Stronger Models
Core Problem: The structural inefficiency in large language models (LLMs) due to 'attention collapse' in deeper layers, leading to redundant 'lazy layers' that hinder model efficiency.
Key Innovation: Identifying the phenomenon of attention collapse and introducing Inheritune, a training recipe that leverages this by inheriting potent early layers from larger models to build smaller, yet stronger and more efficient LLMs.
214. Resource-Efficient Personal Large Language Models Fine-Tuning with Collaborative Edge Computing
Core Problem: The computational intensity and resource scarcity on edge devices hinder the fine-tuning of personal large language models (LLMs) at the network edge, despite the benefits for data privacy and security.
Key Innovation: Proposing Pluto and Charon (PAC), a collaborative edge AI framework that achieves resource-efficient personal LLM fine-tuning through an algorithm-system co-design, including Parallel Adapters, an activation cache, and hybrid data/pipeline parallelism across edge devices.
215. Generation of Uncertainty-Aware High-Level Spatial Concepts in Factorized 3D Scene Graphs via Graph Neural Networks
Core Problem: The manual and heuristic-driven process of generating high-level spatial concepts (like rooms and walls) and their associated optimization factors and covariances within 3D Scene Graphs for robot navigation and mapping, which limits generalization and scalability.
Key Innovation: A novel learning-based method using Graph Neural Networks (GNNs) to autonomously infer uncertainty-aware high-level spatial concepts online from primitive observations and integrate them as optimizable factors within a SLAM backend, improving mapping and localization performance.
216. Are Statistical Methods Obsolete in the Era of Deep Learning? A Study of ODE Inverse Problems
Core Problem: The question of whether traditional statistical methods remain relevant compared to deep learning for modeling, inference, and prediction, particularly in ODE inverse problems with sparse and noisy data.
Key Innovation: Demonstrates through case studies (SEIR, Lorenz models) that statistically principled methods (MAGI) consistently outperform deep learning (PINN) in terms of lower bias/variance, fewer parameters, less hyperparameter tuning, and better out-of-sample prediction for ODE inverse problems, especially with sparse/noisy observations.
217. Inverse Mixed-Integer Programming: Learning Constraints then Objective Functions
Core Problem: Existing data-driven inverse optimization methods primarily focus on learning objective functions under known constraints, leaving the problem of learning both objective functions and constraints from data largely unexplored.
Key Innovation: Proposes a two-stage approach for inverse optimization problems to learn both constraints and objective function weights for mixed-integer linear programs, providing finite-sample guarantees and demonstrating success on scheduling instances.
218. What Can Be Recovered Under Sparse Adversarial Corruption? Assumption-Free Theory for Linear Measurements
Core Problem: Determining what can be uniformly recovered from linear measurements corrupted by a sparse adversarial vector without strong structural assumptions on the matrix or signal, and without knowing the corruption itself.
Key Innovation: Shows that the smallest robust solution set is $x^\star + \ker(U)$ (where U is a specific projection matrix) and proves that minimizing the $\ell_0$-norm of the residual recovers this set, providing an assumption-free theory for recovery under sparse adversarial corruption.
219. Graph Neural Networks for Interferometer Simulations
Core Problem: Simulating complex physical systems, such as the optical physics in interferometers like LIGO, is computationally intensive and time-consuming with state-of-the-art simulation packages.
Key Innovation: Introduces Graph Neural Networks (GNNs) as a new application for instrumentation design, demonstrating their capability to accurately simulate LIGO models while achieving runtimes 815 times faster than traditional methods, and provides a dataset for benchmarking.
220. SlimEdge: Performance and Device Aware Distributed DNN Deployment on Resource-Constrained Edge Hardware
Core Problem: Deploying large distributed deep neural networks (DNNs) on resource-constrained edge devices is challenging due to high computational demands, large parameter counts, and the need for robustness against device failures, while maintaining task performance.
Key Innovation: SlimEdge, an approach that combines structured model pruning with a multi-objective optimization framework to tailor DNN capacity for heterogeneous edge devices. It explicitly accounts for device availability and failure probability, demonstrating improved inference time and maintained accuracy even under multiple device failures, particularly for 3D object recognition.
221. Robust Generalization with Adaptive Optimal Transport Priors for Decision-Focused Learning
Core Problem: Existing Sinkhorn Distributionally Robust Optimization (DRO) methods for few-shot learning rely on fixed reference distributions, limiting their adaptability and robust generalization under distribution shifts.
Key Innovation: Proposes Prototype-Guided Distributionally Robust Optimization (PG-DRO) which learns class-adaptive priors from abundant base data via hierarchical optimal transport, integrating few-shot information for class-specific robust decisions.
222. Predictive Query Language: A Domain-Specific Language for Predictive Modeling on Relational Databases
Core Problem: The manual, slow, and error-prone process of extracting training examples (prediction entities and target labels) from relational databases for machine learning model training.
Key Innovation: Predictive Query Language (PQL), an SQL-inspired declarative language that allows specifying a predictive task in a single query, automating the computation of training labels for a wide variety of machine learning tasks on relational databases.
223. ReaDy-Go: Real-to-Sim Dynamic 3D Gaussian Splatting Simulation for Environment-Specific Visual Navigation with Moving Obstacles
Core Problem: Visual navigation models struggle in real-world dynamic environments due to the sim-to-real gap and difficulty in training policies tailored to target deployment environments, especially with moving obstacles.
Key Innovation: ReaDy-Go, a novel real-to-sim simulation pipeline, synthesizes photorealistic dynamic scenarios by augmenting static 3D Gaussian Splatting (GS) scenes with dynamic human GS obstacles, enabling the training of robust navigation policies for dynamic environments.
224. Improving Regional Forest Biomass Estimation With a UAV-Driven Hierarchical Framework: A Case Study in Chinese Fir Plantations
Core Problem: Accurate regional forest aboveground biomass (AGB) estimation is constrained by the scale mismatch between fine-scale field plots and coarser satellite imagery, particularly in structurally complex forest types like Chinese fir plantations.
Key Innovation: A UAV-driven hierarchical framework integrating a stacking ensemble learning (SEL) strategy that bridges the scale gap by using plot-scale UAV data to generate robust training samples for regional satellite-scale AGB estimation, significantly improving accuracy compared to conventional ground-satellite methods.
225. A Super-Resolution Method for Remote Sensing Images of Coastal Aquaculture Ponds Based on Dynamic Sparse Self-Attention
Core Problem: Precise remote sensing monitoring of coastal aquaculture ponds requires high-resolution imagery, but cost-effective options often lack the necessary detail, hindering effective management.
Key Innovation: A super-resolution method based on a dynamic sparse self-attention mechanism, built on the Transformer architecture, which enhances feature extraction by selectively focusing on relevant image areas and incorporates a scale-aware interaction module to better capture multiscale pond features, outperforming existing methods in image quality and texture preservation.
226. A database of databases for Common Era paleoclimate applications
Core Problem: The challenge of integrating and de-duplicating diverse paleoclimate records from multiple curated databases for comprehensive Common Era (1–2000 A.D.) paleoclimate research.
Key Innovation: Creation of DoD2k version 2.0.0, a merged, duplicate-free database of five curated paleoclimate datasets spanning 13 archive types and 37 data types, along with a toolkit for its creation and management, facilitating comprehensive paleoclimate analysis.
227. FRIDA-Clim v1.0.1: a simple climate model with process-based carbon cycle used in the integrated assessment model FRIDAv2.1
Core Problem: The need for a simple, yet robust and process-based, climate model that can be tightly coupled within an Integrated Assessment Model (IAM) to interactively simulate the co-evolution of social and climate systems, while also serving as a standalone tool for fast climate response calculations.
Key Innovation: Development of FRIDA-Clim v1.0.1, a simple climate model with process-based carbon cycle representations (ocean, land, atmosphere), calibrated to observations, designed for both standalone use and deep integration within the FRIDAv2.1 IAM to study coupled human-Earth system dynamics and explore climate response uncertainty.
228. Modelling framework for asynchronous land-atmosphere coupling using NASA GISS ModelE (NASA-GISS E2.1) and LPJ-LMfire (v1.4.0): design, application and evaluation for the 2.5 ka period
Core Problem: The need for a generalized framework to incorporate biogeophysical responses into climate models without dynamic vegetation for simulating past climates, particularly for the understudied period between the mid-Holocene and the Last Millennium, and to understand the influence of land cover on climate.
Key Innovation: Development and evaluation of an asynchronous land-atmosphere coupling framework using NASA GISS ModelE and LPJ-LMfire for paleoclimate simulations (2.5 ka period), demonstrating strong vegetation-albedo feedback, sensitivity to bias correction, and skill in reproducing past climate and regional precipitation responses.
229. Conservation priority for protected areas in Fuzhou, southeast China: An integrated inside-out approach based on ecological network
Core Problem: Protected areas suffer from internal fragmentation and neglect of surrounding critical habitat networks, leading to conservation gaps and inefficient conservation efforts.
Key Innovation: Introduced the Framework for Conservation Priority Identification (FCPI), integrating MSPA, RSEI, Circuit Theory, and MCR model to formulate a multidimensional conservation priority index, enabling dynamic prioritization and identification of critical ecological network components and conservation gaps.
230. Modelling evacuation dynamics in multi-storey school dormitories under fire conditions
Core Problem: Understanding and improving fire evacuation performance in multi-storey school dormitories, which is challenging due to rising occupancy, complex vertical circulation, and the coupling of pedestrian motion with fire-induced environmental fields.
Key Innovation: Developed an improved social force model coupling pedestrian motion with fire-induced environmental fields (CO, visibility, temperature), calibrated parameters using real-world video observations of adolescent behavior, and systematically examined how fire source location influences evacuation performance, revealing phase-specific spatio-temporal heterogeneity and congestion patterns.
231. Remaining useful life prediction of aero-engine using pyramid temporal convolutional network with fused complementary attention
Core Problem: Existing Remaining Useful Life (RUL) prediction methods for aero-engines struggle with cross-domain and operating condition adaptation, personalized adaptation, and modeling of temporal-spatial coupling characteristics, hindering reliability.
Key Innovation: A novel RUL prediction framework (PTCN-FCA) employing a multi-scale pyramidal temporal convolutional network, personalized parameter adaptation with a lightweight auxiliary branch, a degradation pattern-switching mechanism, and a complementary attention mechanism to enhance accuracy and generalization for aero-engine health management.
232. Deep transfer learning based on cross-domain subsequence alignment and feature contribution interpretation for remaining useful life prediction
Core Problem: Deep transfer learning for Remaining Useful Life (RUL) prediction faces challenges in cross-domain subsequence alignment, robustness under few-shot conditions, and interpretability of model decisions, leading to unstable and unreliable predictions due to significant domain shifts.
Key Innovation: A RUL transfer prediction method that optimizes cross-domain transfer by developing a multi-scale subsequence alignment strategy, constructing an interpretable cross-domain transfer learning model with self-attention and a Bayesian linear layer, and establishing a lightweight analysis method for critical time steps and feature contributions.
233. An interpretable transfer bayesian method for remaining useful life prediction
Core Problem: Accurate and interpretable Remaining Useful Life (RUL) prediction for field equipment remains challenging, especially when dealing with domain shifts and streaming multi-sensor data.
Key Innovation: An interpretable transfer Bayesian method integrating dynamic pseudo-domain generation (DPG) to match degradation trajectories, a dual-scale distance for optimal pseudo-domain unit selection in online Bayesian updating, and adaptive weighting of multi-sensor features to improve RUL prediction accuracy and robustness.
234. Reply to Comment on “Revisiting the Dom Feliciano Belt and surrounding areas – An integrated geophysical and isotope geology approach”
Core Problem: Addressing specific points of contention raised in a comment regarding the interpretation of the Dom Feliciano Belt's tectonic evolution, including terrane terminology, geophysical interpretations, structural continuity, and magmatic arc origins.
Key Innovation: Reaffirming the robustness of the integrated geophysical and geochronological framework for understanding the complex tectonic evolution of southwestern Gondwana, by responding to specific criticisms and maintaining original interpretations.
235. Comment on “Revisiting the Dom Feliciano Belt and surrounding areas – An integrated geophysical and isotope geology approach” by Teixeira et al
Core Problem: The original paper by Teixeira et al. (2025) contains inaccuracies and overlooks relevant publications regarding the Dom Feliciano Belt, leading to potentially flawed interpretations of its tectonic evolution.
Key Innovation: Providing a critical evaluation of the original paper's interpretations, specifically challenging its 'terrenology', the autochthonous nature of tectonic domains, the continuity of shear zones, and the origin of magmatic arcs, to deepen the discussion and ensure a more comprehensive understanding of the region.
236. Surface reflectance in homogeneous regions follows a Gamma distribution: principles, validation and applications
Core Problem: The lack of a rigorous theoretical basis and comprehensive validation for common assumptions (e.g., Gaussian) about the statistical distributions of surface reflectance (SR) in optical remote sensing.
Key Innovation: Reveals and systematically validates that surface reflectance in homogeneous regions follows a Gamma distribution, and its logarithm approximately follows a Gaussian distribution. Integrates this finding into a Bayesian framework to develop the Approximate Bayesian Optimal Classifier (ABOC), which outperforms other machine learning models and global land cover products in classification accuracy.
237. Scale-invariant property of the total source basin area
Core Problem: Understanding the organization of channel networks requires consistent indicators across observational scales, and the explicit mechanistic rationale for previously reported invariant properties of source basins is lacking.
Key Innovation: Demonstrated the scale-invariant property of the total source basin area using a probabilistic approach on DEM data, revealing distinct scaling regimes for ephemeral and perennial networks and establishing this area as a robust geomorphologic indicator for fluvial systems.
238. A parameterization for estimating ice sublimation in the Qilian mountains
Core Problem: Conventional approaches for quantifying ice sublimation, a key component of mass loss from snow/ice surfaces, suffer from considerable uncertainties, making observation and simulation challenging in mountainous regions.
Key Innovation: Established a new empirical formula based on daily maximum temperature, relative humidity, wind speed, and D20 pan sublimation, which significantly outperformed other methods in accurately simulating daily ice surface sublimation in the Qilian Mountains.
239. Large deformation numerical modelling of suction caisson installation in clay: Quantification of soil heave and suction control
Core Problem: Reliable assessment of suction caisson installation in clay, particularly concerning internal soil heave development and suction control, is crucial for design but challenging, with potential misapplication of design principles from jacking installation.
Key Innovation: Used large deformation sequential limit analysis to investigate suction caisson installation, revealing fundamental differences from jacking, negligible influence of tip chamfer, and a significant adverse effect of interface roughness on soil heave and installation refusal, providing explicit empirical models for design.
240. Enhancement of mechanical and carbon sequestration properties of lightweight stabilized soil using ground granulated blast furnace slag, desulfurized gypsum, and CO2-foam
Core Problem: There is a need for sustainable construction materials that can simultaneously achieve lightweighting and permanent carbon dioxide storage, especially when utilizing industrial byproducts and waste slurry.
Key Innovation: Developed Carbon Sequestration Lightweight Soil (CLS) by incorporating CO2 foam into stabilized soils, demonstrating simultaneous enhancement of mechanical strength and reduction in bulk density, high carbon sequestration capacity through carbonate mineralization, and significant environmental potential for infrastructure applications.