TerraMosaic Daily Digest: Feb 19, 2026
Daily Summary
This digest synthesizes 165 selected papers (from 918 deduplicated papers analyzed) spanning landslide mechanics, earthquake source processes, and monitoring workflows that are increasingly built around AI and multi-sensor remote sensing. Two studies on the 2025 Mw 8.8 Kamchatka earthquake exemplify this shift: one uses ETAS and change-point analysis to resolve a multi-decade quiescence-to-activation transition, while the other reconstructs a largely unilateral rupture by combining teleseismic and satellite constraints with deep-ocean tsunami records.
Across slope hazards, the strongest contributions tighten the trigger–process–impact chain. Large-scale loess experiments and acoustic-emission–driven learning frameworks quantify how engineered disturbance, rainfall infiltration, and subsurface deformation evolve toward failure, while susceptibility mapping work couples explainable models with diffusion-based augmentation to better resolve clustered, rainfall-driven landslides under limited labels. Cryosphere papers add a practical early-warning handle by identifying cumulative diurnal freeze–thaw cycles as a precursor to successive glacier collapses, complementing advances in robust change detection and decorrelation-tolerant InSAR aimed at operational monitoring in difficult terrain.
Key Trends
- Quantified precursors are moving from anecdotes to diagnostics: ETAS-based quiescence/activation patterns for megathrust earthquakes and cumulative freeze–thaw counts for glacier-slope failures show how simple, trackable metrics can anchor time-dependent hazard awareness.
- Generative learning is being used selectively to fix the hard parts of mapping: Diffusion-based augmentation targets “confusion zones” in susceptibility mapping, while explainability (e.g., SHAP-style attribution) is treated as part of the deliverable rather than an add-on.
- Monitoring pipelines are being hardened for real-world messiness: Near-real-time polarimetric InSAR for rapid-decorrelation settings, ML interpretation of acoustic emissions, and “hallucination-resistant” multimodal change detection aim to reduce false confidence where data are sparse or noisy.
- Process studies re-center hydro-mechanical coupling and disturbance: Work on loess, dispersive soils, and liquefaction repeatedly highlights how infiltration pathways, pore-pressure evolution, and material fabric govern the transition from stable deformation to runaway failure.
- Engineering relevance is shifting upstream into model design: From tunnel rockburst energy forecasting to reinforced backfills under variable water levels, many papers are framed around decisions (support design, monitoring thresholds, maintenance) rather than post hoc explanation.
Selected Papers
This digest features 165 selected papers from 918 papers analyzed (out of 2611 raw papers scanned; 918 new papers after deduplication). Each paper has been evaluated for its relevance to landslide and broader geohazard research and includes links to the original publications.
1. Seismic quiescence and activation prior to the 2025 M8.8 Kamchatka, Russia earthquake
Core Problem: Understanding the long-term evolution of seismicity preceding major subduction zone ruptures, specifically identifying temporal patterns like multiyear quiescence and short-term activation, is crucial but poorly documented for events like the 2025 M8.8 Kamchatka earthquake.
Key Innovation: Application of the ETAS model and change-point analysis to the 2025 M8.8 Kamchatka earthquake data (1975-2025), revealing a pronounced ~20-year quiescent interval followed by abrupt activation. This finding, consistent with previous megathrust events in the Kamchatka-Kuril system, suggests that heightened seismic activity following multiyear quiescence is a recurrent feature, important for situational awareness and time-dependent hazard assessment.
2. Simple unilateral rupture of the great Mw 8.8 2025 Kamchatka earthquake
Core Problem: Understanding the rupture characteristics, recurrence patterns, and seismic risk of great earthquakes, especially when they deviate from regular patterns.
Key Innovation: Estimation of the slip history for the 2025 Kamchatka earthquake by combining teleseismic and satellite data with deep-water tsunami recordings, revealing complex strain release patterns and heightened seismic risk due to variable recurrence and slip distributions.
3. Multi-field responses and failure mechanisms of loess slopes under engineering disturbance and extreme rainfall: implications for sustainable slope management
Core Problem: Elucidating the stability and failure mechanisms of loess slopes affected by engineering interventions and extreme rainfall is crucial for sustainable slope management in loess terrains, especially in the context of climate change.
Key Innovation: A 1:20 large-scale physical model demonstrated how engineering activities and rainfall interact to cause loess slope failure, detailing significant stress concentrations, spatially and temporally variable rainfall infiltration, pore-water pressure buildup, and a specific failure sequence: 'shallow softening → shallow mud-induced sliding → toe-shear failure → flow-plastic/liquefied sliding'.
4. Landslides impact and management on human settlements over 3000 years: the case of the Montilla Castle Hill (Córdoba, Spain)
Core Problem: Understanding the long-term interactions between human settlements and landslide hazards over millennial timescales, particularly how anthropogenic factors and natural events influence landslide activity.
Key Innovation: Adopted a multidisciplinary approach (archaeology, geology, geophysics, geotechnical modeling) to characterize 3000 years of landslide activity at Montilla Castle Hill, identifying major buildings as a key factor in reactivating ancient landslides and correlating activity with regional earthquakes.
5. Machine learning analysis of landslide subsurface deformation using acoustic emission data
Core Problem: The challenge of quantitatively relating complex and stochastic acoustic emission (AE) signals to actual subsurface landslide deformation, making interpretation difficult for monitoring and early warning.
Key Innovation: Integrated AE monitoring with machine learning (ML) techniques to classify landslide kinematic states and predict displacement with high accuracy, offering a cost-effective, real-time monitoring framework to enhance understanding of landslide dynamics and early warning capabilities.
6. Explainable machine learning and generative diffusion modeling for improved susceptibility mapping of rainfall-induced clustered landslides: A case study from Wuping County, southeastern China
Core Problem: Accurate susceptibility mapping for rainfall-induced clustered landslides is challenging due to strong multi-factor coupling, limited sample conditions, and class boundary ambiguity.
Key Innovation: A hybrid framework combining explainable machine learning (XGBoost with SHAP analysis) and a Denoising Diffusion Probabilistic Model (DDPM) for controlled data augmentation in model 'confusion zones' significantly enhances landslide susceptibility mapping accuracy and interpretability under limited-sample conditions.
7. Spatiotemporal Characteristics of Tectonic Tremors in California
Core Problem: The full spatial distribution of tectonic tremor activity across California and its implications for regional seismic hazard were previously unclear.
Key Innovation: Identified over 83,000 tremor events across California from 2000-2025, revealing new tremor clusters near the Mendocino Triple Junction and within the Big Bend segment, which have implications for regional seismic hazard assessment.
8. Hallucination-Resistant Change Detection in Multimodal Large Models for Autonomous Land Management Agents
Core Problem: Traditional deep learning models for remote sensing change detection lack robustness and generalization, while large language models (LLMs) suffer from 'hallucination' issues, limiting their reliability for critical tasks like landslide mapping.
Key Innovation: Proposes CVAHR-CD-LLM, a change vector analysis (CVA)-based hallucination-resistant multimodal large language model framework. It expands change detection datasets (including landslide mapping) using a semi-supervised approach, introduces a CVA-based coordinations iterative calibration loop, and incorporates a self-consistency mechanism to suppress hallucinations and enhance generalization and accuracy.
9. Freeze–thaw cycles control successive glacier landslide hazards in Amney Machen Mountain
Core Problem: Inadequate understanding of triggering processes for glacier-related landslides in alpine environments, hindering robust hazard assessments and effective risk mitigation.
Key Innovation: Identified extreme diurnal freeze–thaw cycles as the primary driver for glacier collapses and established the 365-day cumulative number of diurnal freeze–thaw cycles as a critical quantitative precursor for initiating these hazards, demonstrating its feasibility for early warning.
10. How trees influence slope stability during cyclones: field and experimental evidence
Core Problem: The complex and underexplored dual role of trees in affecting slope stability during cyclones, where they can both reinforce and destabilize slopes.
Key Innovation: Provided field and experimental evidence demonstrating that tree roots both increase soil porosity/permeability (enhancing infiltration) and provide mechanical reinforcement, highlighting the need to incorporate both stabilizing and destabilizing effects into landslide risk assessments under cyclonic conditions.
11. Post-event investigation of multiscale failure patterns associated with the Mw 6.0 2025 Nurgal-Kunar earthquake, Afghanistan
Core Problem: Understanding the multiscale failure patterns and comprehensive impacts (structural, geotechnical, social, environmental) of a major earthquake in a vulnerable region to inform resilient reconstruction and risk reduction.
Key Innovation: Presented a post-event reconnaissance survey of the 2025 Nurgal-Kunar earthquake, detailing complex failure patterns including slope instabilities and liquefaction, analyzing structural vulnerabilities, and highlighting the severe humanitarian crisis to advocate for resilient reconstruction and improved microzonation.
12. Rheological behavior of dispersive soils: water and cation influence on check dam soils in Northern Shaanxi, China
Core Problem: Understanding the mechanisms behind significant erosion and numerous dam failures caused by widespread dispersive soils in northern Shaanxi, China.
Key Innovation: Uses a rheological approach to investigate the effects of water content, Na+, and Ca2+ on the structural stability and rheological properties of dispersive soils, revealing how these factors influence shear strength and fluidity, and develops a prediction model for rheological properties in plastic stages.
13. Experimental Study on the Rapid Unloading Effect of Rockburst
Core Problem: A systematic understanding of the rapid unloading effect on rockbursts, including macro-mechanical response and meso-scale damage mechanisms, and its dependence on burial depth, is lacking.
Key Innovation: This study revealed distinct nonlinear deformation and acoustic emission characteristics under rapid unloading, identified two key mechanical effects (transient tensile and compressive), and proposed a novel theoretical model incorporating unloading rate to predict damage evolution in rock masses during excavation, directly addressing rockburst mechanisms.
14. Comparison of the shear strain and the deformation test value obtained for the triangular areas affected by Kahramanmaraş earthquakes with a magnitude of Mw 7.7 and Mw 7.6 with the different GNSS networks
Core Problem: Understanding and accurately quantifying surface deformations caused by major earthquakes, and evaluating the impact of different geodetic network configurations on the measurement and interpretation of these deformations.
Key Innovation: Analyzing surface deformations from the Kahramanmaraş earthquakes using static deformation models and 2D strain analyses across three different GNSS networks, demonstrating how network configuration affects deformation test values and identifying areas of maximum shear strain.
15. A review of field monitoring methods for bedrock frost weathering in cold alpine regions
Core Problem: Understanding bedrock frost weathering, a key process driving bedrock degradation, landform evolution, and geological hazards in cold alpine regions, requires robust field monitoring, but current methods and their limitations need systematic review.
Key Innovation: Systematically reviewed the evolution of field monitoring systems for bedrock frost weathering, highlighting the progression from single-temperature measurements to multi-sensor IoT approaches, the expansion of monitored parameters (temperature, moisture, fracture, rockfall/collapse), and discussing current challenges to advance future research and hazard assessment.
16. Data-driven sequential analysis of tipping in high-dimensional complex systems
Core Problem: Quantifying changes in the geometry of attracting sets to detect abrupt transitions ("tipping") in high-dimensional complex systems from partial and noisy observations remains challenging.
Key Innovation: The DA-HASC framework, which combines data assimilation and manifold learning to reconstruct high-dimensional system states and quantify the structural complexity of system dynamics, enabling sequential detection of tipping points relevant to various tipping mechanisms.
17. Physics Encoded Spatial and Temporal Generative Adversarial Network for Tropical Cyclone Image Super-resolution
Core Problem: Existing deep learning-based super-resolution (SR) methods for satellite imagery treat tropical cyclone sequences as generic videos, neglecting the underlying atmospheric physical laws governing cloud motion, leading to meteorologically implausible reconstructions.
Key Innovation: PESTGAN, a Physics Encoded Spatial and Temporal Generative Adversarial Network, which incorporates a PhyCell module to approximate the vorticity equation and encode physical dynamics, along with a dual-discriminator framework for motion consistency and spatial realism, resulting in superior physical fidelity for TC image super-resolution.
18. Adapting Actively on the Fly: Relevance-Guided Online Meta-Learning with Latent Concepts for Geospatial Discovery
Core Problem: Efficiently uncovering hidden targets in dynamic geospatial environments (e.g., environmental monitoring, disaster response) is challenging due to costly data collection, sparse ground truth, and dynamic conditions, limiting existing learning methods.
Key Innovation: Proposes a unified geospatial discovery framework integrating active learning, online meta-learning, and concept-guided reasoning, featuring a concept-weighted uncertainty sampling strategy and a relevance-aware meta-batch formation strategy to improve efficiency and generalization in dynamic environments.
19. Point-DeepONet: Predicting Nonlinear Fields on Non-Parametric Geometries under Variable Load Conditions
Core Problem: Nonlinear structural analyses in engineering require extensive finite element simulations, and conventional deep learning surrogates struggle with complex, non-parametric 3D geometries and directionally varying loads, limiting real-time application and design optimization.
Key Innovation: Presents Point-DeepONet, an operator-learning-based surrogate that integrates PointNet into the DeepONet framework to learn a mapping from raw point clouds of non-parametric geometries and variable load conditions to 3D displacement and von Mises stress fields, achieving high fidelity and significantly faster predictions compared to FEM.
20. TriPhysGAN-Attn: A Physics-Informed Generative Model for Radar Echo Forecasting via Triple Mechanism Decomposition and Attention Fusion
Core Problem: Existing deep learning-based radar echo prediction models lack clear physical mechanisms and struggle to represent the multiscale dynamic processes of precipitation, making them less effective in handling severe convective weather for nowcasting.
Key Innovation: Proposes TriPhysGAN-Attn, a physics-informed generative model that decomposes precipitation evolution into three fundamental processes (Advection, initiation/dissipation, deformation) with dedicated branches. It incorporates a physically informed loss function, a multihead cross-attention mechanism for dynamic information fusion, and a generative adversarial framework with a coordinate attention module.
21. Towards an operational European Drought Impacts Database (EDID)
Core Problem: Lack of a generally accepted convention and operational database for drought impact data collection and use, hindering proactive drought management, robust risk assessment, and development of sustainable adaptation strategies in Europe.
Key Innovation: Development and implementation of the European Drought Impact Database (EDID) for operational application, integrating regional datasets to provide comprehensive, spatially and temporally resolved drought impact information, thereby supporting drought risk management and policy for a drought resilient society.
22. Prepare for extreme dust storms
Core Problem: The increasing threat of extreme dust storms necessitates effective preparedness strategies to mitigate their impacts.
Key Innovation: The paper likely discusses the importance of or methods for preparing for extreme dust storms, implying advancements in forecasting, readiness, and response protocols (details are inferred from the title due to a placeholder abstract).
23. Rapid post-landslide vegetation regrowth detected by multi-temporal satellite imagery in the Southern part of Mt. Rinjani National Park, Lombok, Indonesia
Core Problem: Limited understanding of vegetation recovery periods after landslides in tropical environments and the factors affecting this recovery, which is crucial for assessing re-establishment of slope stability.
Key Innovation: Estimated that full vegetation recovery after landslides in a tropical environment (Mt. Rinjani) takes approximately 6 years, shorter than non-tropical regions, using multi-temporal satellite imagery (NDVI) and identified elevation as the most contributing factor.
24. Influence of water level on seismic behavior of caisson quay walls with fiber-reinforced calcareous sand backfill
Core Problem: Caisson quay walls with calcareous sand backfill are prone to large deformation or instability during earthquakes, especially under large water level fluctuations, and the seismic behavior of fiber-reinforced versions under these conditions has not been systematically studied.
Key Innovation: Shaking table tests demonstrated that fiber reinforcement effectively limits excess pore water pressure below the liquefaction threshold, even under high water levels and strong seismic excitation, significantly enhancing the seismic performance and stability of caisson quay walls.
25. Block Elbowing and Rotation Patterns in Chains of Rounded Blocks
Core Problem: Understanding block motions in blocky rock masses, particularly the influence of rounded block corners and elbowing interactions, is crucial for accurate modeling, as current models may not fully capture these realistic block shapes and their kinematics.
Key Innovation: Discrete element method (DEM) modeling of rounded blocks revealed a transition from reversible to irreversible passive block kinematics governed by contact friction and corner geometry, identifying a previously unreported rotation pattern characterized by residual rotations for highly rounded blocks in short chains, thus improving the understanding of block motion in rock masses.
26. Evolution of Stress-Induced Damage and Anisotropy Prior to Rock Failure in the Main Lithologies of El Teniente Mine, Central Chile
Core Problem: Understanding rock deformation and seismic behavior under deep mining conditions and elevated stress prior to failure requires a holistic characterization of stress-induced anisotropy.
Key Innovation: Triaxial tests with simultaneous acoustic emissions and active seismic surveys on various lithologies revealed distinct mechanical and seismic responses, showing directionally dependent changes in P-wave velocity (Vp) at significant stress levels prior to failure, which may aid in the design of predictive monitoring strategies in deep mining environments.
27. Evolution of Characteristic Stresses in Deep Hard Rocks Under True Triaxial Stresses: The TT-CVS Identification Approach
Core Problem: Existing characteristic stress calculation methods are inadequate for deep rock under true triaxial conditions, failing to account for the significant effect of intermediate principal stress (σ2).
Key Innovation: A novel TT-CVS (True Triaxial-Crack Volumetric Strain) method was proposed to identify rock characteristic stresses under true triaxial conditions, accurately decoupling strain and accounting for the regulatory role of σ2, providing new insights into the evolution laws of characteristic stresses and guiding understanding of brittle disaster mechanisms in deep hard rock.
28. Dynamic Behaviour of Initially Anisotropic Sand under Hydro-Mechanical Coupling based on Resonant Column Tests
Core Problem: Existing studies on the dynamic properties of sand predominantly focus on isotropic saturated sands, with limited research on initially anisotropic sands at varying moisture contents under hydro-mechanical coupling.
Key Innovation: Resonant column tests on reconstituted standard sand analyzed the coupled effects of moisture content, initial anisotropy, and effective confining stress on dynamic shear modulus and damping ratio, proposing modifications to the Hardin and Davidenkov models to predict dynamic shear modulus for initially anisotropic sands, which is critical for assessing liquefaction potential.
29. A physics-informed deep learning and probabilistic inference framework for real-time single-station earthquake detection and magnitude estimation
Core Problem: Established Earthquake Early Warning (EEW) systems require dense seismic networks, handcrafted features, or multi-station triangulation, making them costly and unsuitable for resource-constrained areas.
Key Innovation: Developing a hybrid single-station framework that uses physics-informed preprocessing, a U-Net++ encoder-decoder with MHSA, and Bayesian MCMC for real-time P-wave detection and probabilistic magnitude estimation, achieving significant improvements in F1-scores and reduced magnitude-estimation errors.
30. A building exposure model for Java, Indonesia, for use in Seismic Risk Assessment
Core Problem: A limited understanding of the spatial distribution of building types and the absence of a robust, multi-use, village-scale building exposure model for Java, Indonesia, hinders accurate seismic risk assessment and disaster management policy.
Key Innovation: Developing a desa (village) scale building exposure model for Java by integrating open-source, government, and Census data, identifying sources of uncertainty, and proposing a technique to account for classification uncertainty in masonry residential houses, resulting in a valuable tool for seismic risk assessments.
31. Non-ergodic ground motion model using small-magnitude ground motion data for a site-specific PSHA in Slovenia
Core Problem: Seismic risk assessment can be biased by the ergodic assumption in Ground-Motion Models (GMMs) used in Probabilistic Seismic Hazard Analysis (PSHA), particularly due to the lack of local strong ground-motion data.
Key Innovation: Introducing a methodology for developing a non-ergodic GMM for site-specific PSHA using local small-magnitude ground motion data, involving Bayesian Gaussian process regression and Markov chain Monte Carlo, which results in reduced aleatory standard deviation and a more accurate hazard curve.
32. Static liquefaction: The role of grain size polydispersity from a micro-structural perspective
Core Problem: The triggering mechanisms of static liquefaction at the particle scale are poorly understood, despite its role in catastrophic failures of earthfills, waste dumps, and tailings storage facilities.
Key Innovation: Numerical simulations using the Discrete Element Method reveal a dual micro-structural mechanism for static liquefaction: a collapse of the contact network and the emergence of low-density regions, differentiating between temporary and full liquefaction based on the permanence of these microstructural changes.
33. System reliability-aided assessment of reinforced soil retaining wall stability under static loading using ELM and LSSVM models
Core Problem: Efficient and accurate assessment of component and system reliability for reinforced soil (RS) retaining walls under static loading conditions, considering multiple critical limit states and the computational intensity of traditional probabilistic methods.
Key Innovation: Integrated Extreme Learning Machine (ELM) and Least-Squares Support Vector Machine (LSSVM) models with a sequential compounding method (SCM) to evaluate system reliability of RS retaining walls, demonstrating LSSVM's superior predictive accuracy and computational efficiency as a surrogate for probabilistic analysis, and identifying bearing capacity as the governing failure mode.
34. Seventy‐Five Years Underestimating Frequent Events and Other Frequently Underestimated Implications of Langbein's Equation
Core Problem: Misconceptions and underestimation in the statistical analysis of frequent geophysical events, particularly when comparing annual maxima versus partial duration series methods.
Key Innovation: Clarifying and addressing the implications of Langbein's equation for adjusting annual-maxima-based predictions to improve the accurate estimation of frequent geophysical events.
35. Three-dimensional Damage Visualization of Civil Structures via Gaussian Splatting-enabled Digital Twins
Core Problem: Precise three-dimensional (3D) damage visualization on digital twins of civil structures is needed, transcending traditional 2D image-based methods and conventional photogrammetric 3D reconstruction, especially for post-earthquake inspections.
Key Innovation: Introduced a Gaussian Splatting (GS)-enabled digital twin method for effective 3D damage visualization, which utilizes GS-based reconstruction to visualize 2D damage segmentation results, reduces errors, employs a multi-scale strategy, and enables dynamic updates as damage evolves.
36. TopoFlow: Physics-guided Neural Networks for high-resolution air quality prediction
Core Problem: The need for efficient and accurate high-resolution air quality prediction, which requires explicitly embedding critical physical processes like topography and wind direction into learning frameworks.
Key Innovation: TopoFlow, a physics-guided neural network that integrates topography-aware attention and wind-guided patch reordering into a vision transformer architecture, explicitly modeling terrain-induced flow patterns and wind-driven transport, achieving significant improvements in high-resolution air quality prediction.
37. i-PhysGaussian: Implicit Physical Simulation for 3D Gaussian Splatting
Core Problem: Current 3D reconstruction-based physical simulators rely on explicit, step-wise updates, which are sensitive to step time and suffer from rapid accuracy degradation under complicated scenarios like high-stiffness materials or quasi-static movement, hindering robust risk management.
Key Innovation: i-PhysGaussian, a framework that couples 3D Gaussian Splatting (3DGS) with an implicit Material Point Method (MPM) integrator, which obtains an end-of-step state by minimizing a momentum-balance residual through implicit Newton-type optimization, significantly reducing time-step sensitivity and ensuring physical consistency.
38. Flickering Multi-Armed Bandits
Core Problem: Traditional Multi-Armed Bandit (MAB) frameworks do not account for scenarios where the set of available actions (arms) can change dynamically and depend on previous choices, which is relevant for exploration in constrained environments.
Key Innovation: Introducing Flickering Multi-Armed Bandits (FMAB), a new MAB framework for dynamically changing arm availability, and proposing a two-phase algorithm (lazy random walk for exploration, navigation/commitment for exploitation) with sublinear regret bounds, demonstrated in a robotic disaster scouting scenario.
39. Enhanced Deep Recurrent Optical Flow With Efficient Feature Encoding and Channel Attention for River Surface Velocimetry
Core Problem: Uncrewed aerial vehicle (UAV) image velocimetry for river-surface flow velocity estimation suffers from degraded robustness and accuracy under complex conditions (illumination variability, low-texture surfaces, nonrigid motion), particularly in detecting subtle surface displacements in low-texture regions, which affects flow-field reconstruction and flood early-warning systems.
Key Innovation: Proposing RSV-RAFT, a deep optical-flow framework for river-surface velocity assessment. It enhances sensitivity to weak textures and small-scale motion, designs a decoupled enhanced channel-attention flow head for minute displacements, and integrates a deformable coordinate-aware upsampling mask generator for improved boundary perception, outperforming existing methods in accuracy and efficiency for flood early-warning systems.
40. Effective Frequency Band Driven Limited-Sample Deep Learning High-Resolution Seismic Processing Method
Core Problem: Deep learning high-resolution seismic processing methods face challenges in achieving high-signal-to-noise ratio (SNR) results for field seismic data due to limited effective frequency band, limited training samples, and generalization performance issues.
Key Innovation: Proposes ELDHM, an effective frequency band driven limited-sample deep learning high-resolution seismic processing method. It comprises ETG (synthetic seismic training sample generation with matched effective frequency band), DHN (deep learning network for high-SNR high-resolution output), and ESO (optimization to minimize frequency band impacts and preserve phase).
41. TF-RoadNet: A Topo-Tree Scan and Frequency-Aware Network for Road Extraction From Remote Sensing Images
Core Problem: Accurate road extraction from remote sensing imagery is hindered by severe occlusions, complex road structures, and background clutter, making it difficult to preserve road topological connectivity and enhance edge clarity.
Key Innovation: Proposes TF-RoadNet, a lightweight architecture integrating a topo-tree mamba (models road topologies using a topo-tree scan and dynamically constructed minimum spanning tree) and a frequency-aware enhancement module (improves boundary accuracy by decoupling and enhancing low/high-frequency components using discrete wavelet transform). Also introduces a new complex-occlusion (CO) Road dataset.
42. A Continuous Perspective-Based Strategy for Deep Learning in Semantic Segmentation for Remote Sensing Fault Mapping
Core Problem: Traditional automatic extraction of linear geological faults from remote sensing imagery suffers from low accuracy, discontinuity, and reliance on manual thresholding, which are critical for earthquake studies and disaster prevention.
Key Innovation: Proposes a continuous perspective-based deep learning strategy using a U-Net-like architecture with center-masked loss and multiscale attention to simulate expert multiview observation. It also includes a postprocessing module with ensemble inference and Zhang-Suen thinning to generate continuous linear outputs without manual thresholds, significantly improving recall and fault localization accuracy.
43. Characteristics of human-induced forest fires in China by integrating information from court sentencing records and multiple geospatial data
Core Problem: Efficient forest management policies require a comprehensive understanding of the factors and characteristics behind human-induced forest fire ignitions, particularly in regions like China where such ignitions can lead to legal convictions.
Key Innovation: Integrates information from court sentencing records with multiple geospatial data to characterize human-induced forest fires in China, providing a novel approach to understand ignition patterns and inform forest management strategies.
44. Living on top of a volcanic-hydrothermal system: the case of Copahue village, Argentina
Core Problem: The multifaceted and underreported hazards posed by volcanic-hydrothermal gas emissions and extreme ambient temperatures to a village built directly on such a system, leading to structural damage and health risks.
Key Innovation: Conducted the first gas hazard assessment for Copahue village using mixed methods, revealing direct relationships between building damage and degassing areas, exceeding safety limits for indoor CO2/H2S, and presenting the first CO2 susceptibility map for land-planning and risk reduction.
45. A novel method for gob-side entry retaining in retracement channel: A case study
Core Problem: Addressing the issue of repetitive production system deployment and associated ground stability challenges in irregular mining areas under longwall mining.
Key Innovation: Proposes a novel method for gob-side entry retaining in a retracement channel, investigating its retaining mechanism, developing a roof mechanical model to analyze deformation, and validating a control technology through numerical simulation and field tests, achieving an integrated mining-retreating-retaining process.
46. Factors influencing the effectiveness of acid sulfate soil stabilization with cement-based binders
Core Problem: Sulfide-rich soils (acid sulfate soils) pose significant challenges for construction due to their low bearing capacity, shear strength, and acidification potential.
Key Innovation: Compares the effectiveness of Portland cement and Multicem binders in stabilizing acid sulfate soils, analyzing UCS, porosity, pH, and strain behavior, demonstrating how binder type, dosage, initial pH, and soil composition significantly influence stabilization performance, and offering guidance for optimizing binder selection.
47. Grain size-dependent damage evolution in granite during triaxial compression and confining pressure unloading: acoustic emission characterization and constitutive modeling
Core Problem: Insufficient understanding of the role of grain size in damage evolution of granite during confining pressure unloading, which is critical for assessing rock mass stability in deep underground excavations.
Key Innovation: Integrating triaxial testing and acoustic emission monitoring to characterize grain size-dependent damage mechanisms in granite under unloading conditions, revealing distinct damage evolution and fracture patterns, and proposing a grain size-dependent damage constitutive model incorporating AE energy accumulation.
48. A multi-scale assessment of the effects of runoff/sediment discharge in karst catchments as shown by vegetation-related remote sensing indicators
Core Problem: Few studies have investigated which vegetation-related remote sensing indicator can best capture the variability in runoff and sediment discharge at multiple time scales, especially in ecologically fragile karst regions.
Key Innovation: Identifies Solar-Induced chlorophyll Fluorescence (SIF) as the most effective explanatory variable for both river runoff and sediment in global karst areas, providing new insights for water and soil resource management.
49. Three-dimensional identification and assessment of compound droughts from a propagation perspective
Core Problem: Existing studies on compound droughts primarily focus on temporal overlaps, neglecting the crucial spatiotemporal continuity and propagation of drought evolution, making their identification and assessment challenging.
Key Innovation: A novel dual spatiotemporal coupling framework that uses three-dimensional clustering to identify spatiotemporally continuous drought events and then performs event-level spatiotemporal coupling across drought types from a propagation perspective, enabling effective identification and systematic evaluation of compound droughts and their drivers.
50. Integrating material–scale freezing point determination with ERA5–land reanalysis data for physics–based freeze–thaw zoning along national highway G214 on the Qinghai–Tibet plateau
Core Problem: There is a disconnection between material properties (e.g., concrete freezing point) and regional freeze-thaw zoning, hindering accurate design and long-term maintenance of infrastructure in cold regions like the Qinghai-Tibet plateau.
Key Innovation: Integrated laboratory-determined concrete freezing point data with corrected ERA5-Land reanalysis data to establish physics-based freeze-thaw zoning along National Highway G214, quantifying annual freeze-thaw cycles (NFTCs) using a material-specific threshold, thereby enhancing the scientific rigor for frost-resistant design of concrete structures.
51. Evaluation of tunnel face stability using slip curves based on lower bound theory
Core Problem: Conventional methods for evaluating tunnel face stability rely on unrealistic assumptions of vertical slip lines and an undetermined lateral earth pressure coefficient, particularly in loose ground conditions.
Key Innovation: A novel method based on lower-bound theory is proposed, using curved slip lines modeled as a logarithmic spiral and introducing a stress-state relationship on the slip line to eliminate the need for a lateral earth pressure coefficient, accurately predicting face stability indices.
52. Stress and Deformation Characteristics of the Central‐Southern Tibetan Crust: Insights From Electrical Anisotropy Studies
Core Problem: Mechanisms governing stress-induced surface deformation in continental collision zones like the Tibetan Plateau, particularly the role of crustal rheological properties, remain vigorously debated.
Key Innovation: Electrical anisotropy studies reveal distinct N-S anisotropy in the middle crust (fluid-filled fractures, brittle deformation) and E-W anisotropy in the lower crust (deformed melt pockets, ductile deformation), providing a novel geo-electrical method to characterize crustal deformation and weakening in active regions.
53. Regionalization of Hydrologic Behavior and Pothole Water Storage Dynamics in the Prairie Pothole Region
Core Problem: Challenges in predicting streamflow and pothole water storage dynamics in ungauged, pothole-dominated catchments due to complex fill-spill-connection mechanisms and data scarcity, hindering vulnerability and flood-risk assessment.
Key Innovation: Development of δHBV-Pot, a physics-informed deep learning model integrating conceptual hydrology with a probabilistic algorithm to emulate pothole fill-spill-connection processes, enabling regionalization of high-flow and pothole storage characteristics for improved flood-risk assessment in ungauged catchments.
54. StereoAdapter-2: Globally Structure-Consistent Underwater Stereo Depth Estimation
Core Problem: Stereo depth estimation in underwater environments suffers from severe domain shifts (light attenuation, scattering, refraction), and existing GRU-based iterative refinement methods struggle with long-range disparity propagation, limiting performance in large-disparity and textureless regions.
Key Innovation: StereoAdapter-2, which replaces the conventional ConvGRU updater with a novel ConvSS2D operator based on selective state space models, enabling efficient, globally structure-consistent long-range spatial propagation within a single update step. Also, introduces UW-StereoDepth-80K, a large-scale synthetic underwater stereo dataset.
55. HS-3D-NeRF: 3D Surface and Hyperspectral Reconstruction From Stationary Hyperspectral Images Using Multi-Channel NeRFs
Core Problem: Integrating hyperspectral imaging (HSI) and 3D reconstruction at scale is challenging, as conventional approaches involve complex hardware setups or require moving-camera setups, limiting throughput and reproducibility in controlled environments.
Key Innovation: HSI-SC-NeRF, a stationary-camera multi-channel NeRF framework for high-throughput hyperspectral 3D reconstruction. It uses multi-view hyperspectral data from a stationary camera with object rotation, ArUco marker-based pose estimation, and a multi-channel NeRF formulation with a composite spectral loss and two-stage training.
56. Synergizing Transport-Based Generative Models and Latent Geometry for Stochastic Closure Modeling
Core Problem: Diffusion models, while effective for generative AI, suffer from slow sampling speeds, which is a disadvantage for learning stochastic closure models in complex dynamical systems, and ensuring physical fidelity in latent space is crucial.
Key Innovation: Demonstrating that flow matching in a lower-dimensional latent space enables single-step, faster sampling for stochastic closure models, and introducing explicit (metric-preserving, geometry-aware) and implicit regularization to control latent space distortion and ensure physical fidelity, requiring less training data.
57. TIFO: Time-Invariant Frequency Operator for Stationarity-Aware Representation Learning in Time Series
Core Problem: Nonstationary time series forecasting suffers from the distribution shift issue between training and test data, and existing methods fail to capture the underlying time-evolving structure across samples or model complex time structures effectively.
Key Innovation: TIFO (Time-Invariant Frequency Operator), which learns stationarity-aware weights over the frequency spectrum across the entire dataset, highlighting stationary frequency components while suppressing non-stationary ones, thereby mitigating the distribution shift issue and improving forecasting accuracy and computational efficiency.
58. 3D Scene Rendering with Multimodal Gaussian Splatting
Core Problem: Conventional vision-based 3D Gaussian Splatting (GS) pipelines rely on a sufficient number of camera views and perform poorly in conditions where visual cues are unreliable (e.g., adverse weather, low illumination, partial occlusions), incurring additional processing costs.
Key Innovation: A multimodal framework that integrates RF sensing (e.g., automotive radar) with GS-based rendering, enabling efficient depth prediction from sparse RF-based depth measurements to yield high-quality 3D point clouds for robust initialization of Gaussian functions across diverse GS architectures, achieving high-fidelity 3D scene rendering in challenging conditions.
59. Inferring Height from Earth Embeddings: First insights using Google AlphaEarth
Core Problem: Effectively translating geospatial and multimodal features encoded in Earth Embeddings into accurate regional surface height estimates using deep learning regression models, particularly concerning generalization to areas with distribution shifts.
Key Innovation: Demonstrates that AlphaEarth Embeddings, when combined with U-Net and U-Net++ architectures, can effectively guide deep learning models for surface height mapping, achieving strong training performance and showing promising potential for regional transferability, despite challenges in generalization.
60. Tree crop mapping of South America reveals links to deforestation and conservation
Core Problem: Monitoring tree crop expansion for zero-deforestation policies is hampered by a lack of high-resolution data that accurately distinguishes diverse agricultural systems from forests, leading to misclassification and potential false alerts.
Key Innovation: The paper presents the first 10m-resolution tree crop map for South America, generated using a multi-modal, spatio-temporal deep learning model on Sentinel-1 and Sentinel-2 data, providing a high-resolution baseline to improve deforestation monitoring and conservation policies.
61. A High-Level Survey of Optical Remote Sensing
Core Problem: The vast and diverse literature on optical remote sensing lacks a comprehensive, high-level overview to guide new researchers and provide holistic insights into the field's capabilities, tasks, and methodologies.
Key Innovation: This paper provides a comprehensive, high-level survey of optical remote sensing, encompassing diverse tasks, capabilities, and methodologies, along with key information like datasets and insights, serving as a guide for researchers entering the field.
62. OpenEarthAgent: A Unified Framework for Tool-Augmented Geospatial Agents
Core Problem: Extending multimodal reasoning capabilities to the remote sensing domain is challenging, as models must reason over spatial scale, geographic structures, and multispectral indices while maintaining coherent multi-step logic for analytical tasks.
Key Innovation: OpenEarthAgent, a unified framework for developing tool-augmented geospatial agents trained on satellite imagery, natural-language queries, and detailed reasoning traces. It uses supervised fine-tuning over structured reasoning trajectories and includes a large corpus spanning urban, environmental, disaster, and infrastructure domains, demonstrating structured reasoning and stable spatial understanding.
63. Neural Implicit Representations for 3D Synthetic Aperture Radar Imaging
Core Problem: 3D Synthetic Aperture Radar (SAR) imaging suffers from significant artifacts in reconstructed imagery due to sparse data in the Fourier domain, making accurate representation of complex scenes challenging.
Key Innovation: Employs neural implicit representations to model surface scattering, encoding object surfaces as signed distance functions learned from sparse SAR data, and regularizing surface estimation by sampling points from this implicit representation during training to achieve state-of-the-art results.
64. Risk-Aware Decision Making in Restless Bandits: Theory and Algorithms for Planning and Learning
Core Problem: Traditional restless bandits problems use a risk-neutral objective, which is insufficient for real-world applications where mitigating downside risks is crucial.
Key Innovation: Generalizes the restless bandits problem to incorporate risk-awareness, establishing indexability conditions and providing a Whittle index solution for planning, and proposing a Thompson sampling approach for learning with unknown transition probabilities, demonstrating efficacy in reducing risk exposure.
65. Toward Cross-Regional Land Cover Mapping Using Regional Consistency and Certainty-Based Active Domain Adaptation With Limited Annotations
Core Problem: In cross-domain semantic segmentation for land cover mapping using remote sensing images, accurately identifying boundaries between different types of surface features remains a critical challenge, especially with limited annotations when transferring models across regions.
Key Innovation: Introducing a region-aware active selection strategy incorporating regional consistency and certainty for remote sensing segmentation, called cross-regional active land cover mapping. This method enhances feature discriminability by maintaining local prediction consistency and certainty, leading to more accurate boundary identification and significant improvement in segmentation accuracy with minimal annotation effort.
66. A Real-Time MPSoC-Based Back Projection Accelerator for High-Accuracy Large-Size SAR Imaging Using Truncated Sinc Reconstruction and Mixed Precision Design
Core Problem: The back projection (BP) algorithm for high-resolution SAR imaging has high computational complexity, severely limiting real-time performance and scalability, especially for large-size, high-accuracy imaging on resource-constrained systems.
Key Innovation: Proposes an efficient BP acceleration architecture based on truncated sinc interpolation reconstruction, which reduces off-chip memory bandwidth, on-chip memory, and logic resource consumption. It also introduces a mixed precision strategy to reduce lookup table consumption while maintaining imaging accuracy, achieving real-time processing for large-size, high-accuracy SAR imaging.
67. Highly scalable geodynamic simulations with HyTeG
Core Problem: The computational challenge of performing high-resolution global geodynamic simulations (e.g., mantle convection) due to prohibitively high memory demands and low arithmetic intensity of traditional sparse matrix methods.
Key Innovation: Leveraging the matrix-free Finite Element framework HyTeG to conduct highly scalable geodynamic simulations with realistic physical models, demonstrating excellent performance up to 100 billion unknowns and validating its accuracy against analytical solutions and geophysical benchmarks.
68. Catchment transit time variability with different SAS function parameterizations for the unsaturated zone and groundwater
Core Problem: Uncertainty in representing preferential flow paths and their impact on catchment transit time distributions (TTD) using isotope-based transport models, particularly in distinguishing the influence of the unsaturated zone versus groundwater aquifers on streamflow tracer signals.
Key Innovation: Systematic comparison of different StorAge Selection (SAS) function parameterizations for the unsaturated zone and groundwater, demonstrating that streamflow δ2H measurements can isolate preferential flow in the unsaturated zone but are insufficient to constrain preferential groundwater flow in catchments with large passive groundwater storage, highlighting the need for complementary data to reduce TTD uncertainty.
69. MVarGOSIM: an MPS algorithm for characterizing complex structures with multiple variables
Core Problem: Creating highly accurate large-scale geological models is challenging due to sparse data, and effectively incorporating and amalgamating patterns from diverse variables during the simulation process remains a significant obstacle in pattern-based methodologies.
Key Innovation: Presents MVarGOSIM, a novel iterative multiple-point statistics (MPS) algorithm that integrates co-located multiple variables (e.g., lithology, velocity, density) using fully connected deep artificial neural networks (FCNs) to construct complex geological structures, demonstrating improved accuracy in reconstructing geological objects and preserving their geometry and interrelationships.
70. Understanding atmospheric dynamics of the 2024 Arabian Gulf flood: Can extreme rainfall be predicted weeks ahead?
Core Problem: Understanding the atmospheric dynamics driving extreme rainfall events and improving their predictability at subseasonal lead times to enhance disaster preparedness.
Key Innovation: Investigated the atmospheric mechanisms behind the 2024 Arabian Gulf flood and demonstrated that subseasonal ensemble forecasts at convection-permitting resolution could provide probabilistic guidance for heavy rainfall two weeks in advance, advancing early detection capabilities.
71. Dilatancy behavior of marine coral sand–fines mixture in the South China Sea: critical thresholds and particle breakage effects
Core Problem: Accurately understanding the dilatancy of marine coral sand containing fines is critical for ensuring the safety of ocean infrastructures, but existing models have limitations in describing this behavior, especially considering particle breakage.
Key Innovation: Experimentally investigated the dilatancy of marine coral sand-fines mixtures, identified critical thresholds related to fines content and void ratio, established relationships between dilatancy ratio and stress/equivalent void ratio, and proposed a semi-empirical dilatancy equation accounting for particle breakage.
72. Simplified Analytical Method for Exploring the Life-Cycle Safety of Tunnels Using a Composite Yielding Support System
Core Problem: Preventing lining damage in deep soft rock tunnels due to time-dependent deformation of the surrounding rock throughout the entire life cycle requires a robust support system, and a theoretical model for composite yielding support systems is needed.
Key Innovation: A theoretical model and analytical solutions were developed to predict the life-cycle mechanical responses of tunnels employing a composite yielding support system, considering three-stage deformation characteristics of the primary lining and buffer layer, providing practical design suggestions for tunnel safety.
73. Effects of the inherent damping modelling approaches on the seismic demand of nonstructural elements
Core Problem: Existing building code recommendations for inherent damping modeling in seismic analysis primarily focus on structural response, neglecting their impact on nonstructural seismic demand, leading to potential miscalculations and erroneous evaluations.
Key Innovation: Evaluating the significant impact of various inherent structural damping modeling assumptions (mass/stiffness proportional, Rayleigh, initial/tangent stiffness, damping levels) on the seismic demand of nonstructural elements, showing mean differences up to three times the minimum response.
74. Neural network based on high freedom dynamic collaboration of flows for meteorological data temporal downscaling
Core Problem: Temporal refinement of meteorological fields is challenging, as conventional approaches are computationally intensive or exhibit limited generalization when applied to complex atmospheric conditions.
Key Innovation: Proposes a deep neural network that independently estimates spatially adaptive kernel parameters for each variable and dynamically integrates information across multiple meteorological fields, achieving high accuracy, robustness, and efficiency for temporal downscaling.
75. Estimation of the monthly change in soil water storage in two watersheds of contrasting vegetation cover
Core Problem: The challenge of accurately estimating catchment-wide monthly changes in soil water storage (ΔW), especially in watersheds undergoing vegetation restoration, and understanding the impact of different vegetation types on these dynamics.
Key Innovation: Integration of the Generalized Complementary Relationship (GCR) of evapotranspiration into a water balance model to accurately estimate ΔW and ETa in watersheds with contrasting vegetation covers, providing insights into the hydrological impacts of plantation vs. grassland restoration on the Loess Plateau and offering suggestions for improved restoration strategies.
76. Soil hydraulic pedotransfer functions for estimating saturated hydraulic conductivity: a deep symbolic regression approach with mean shift clustering
Core Problem: The challenge of accurately estimating soil saturated hydraulic conductivity (Ks) while balancing predictive accuracy, interpretability, and integrability of pedotransfer functions (PTFs) into process-based models.
Key Innovation: A novel two-step framework, Deep Symbolic Regression coupled with Mean Shift clustering (DSR-MS), which partitions soil data into homogeneous clusters and then applies symbolic regression to derive explicit, interpretable mathematical formulas for Ks, achieving high accuracy and integrability for Earth system models.
77. Response analysis of concrete gravity dams based on direct numerical simulation method for wave propagation in reservoir water
Core Problem: Accurately simulating seismic wave transmission through impounded water and its effects on concrete gravity dams is challenging, as indirect simulation methods can lead to significant errors and underestimation of dam damage.
Key Innovation: A direct numerical simulation method for wave transmission through impounded water is proposed, which accurately matches analytical solutions and reveals that indirect simulation significantly underestimates wave energy propagation into the reservoir, leading to underestimated dam stresses and damage.
78. Chemical‐Thermomechanical Modeling of Open‐System Mass Transfer: Application to the Subduction Interface
Core Problem: Simulating subsolidus metasomatism is limited by the decoupling between thermomechanical and chemical/thermodynamic models, leading to high uncertainty in mass transport predictions at subduction interfaces.
Key Innovation: A new fluid-rock chemical-thermomechanical interaction scheme integrates thermodynamic modeling into a 2D thermomechanical code, enabling tracking of slab-derived fluid fluxes and heterogeneous geochemical signatures, showing element precipitation at chemical contrasts and dominant solid-rock advective transport.
79. Structured Analytic Mappings for Point Set Registration
Core Problem: The need for accurate and efficient non-rigid point set registration, particularly for small and smooth deformations, without relying on complex kernel functions or high-dimensional parameterizations.
Key Innovation: An analytic approximation model for non-rigid point set registration based on multivariate Taylor expansion, leading to 'Analytic-ICP,' an efficient algorithm that unifies rigid, affine, and nonlinear deformations under a single closed-form formulation, demonstrating higher accuracy and faster convergence than classical methods.
80. Position: Why a Dynamical Systems Perspective is Needed to Advance Time Series Modeling
Core Problem: Current time series (TS) modeling, despite advancements, lacks a fundamental theoretical framework from dynamical systems (DS) theory, limiting its ability to provide optimal forecasts, predict long-term statistics, generalize to unseen regimes (like tipping points), or inform control strategies.
Key Innovation: Argues for integrating a dynamical systems (DS) perspective into TS modeling, advocating for DS reconstruction (DSR) to infer underlying DS models from data, which would enable theoretically optimal forecasts, prediction of long-term system statistics, generalization to unseen regimes (e.g., tipping points), and inform control strategies, thereby advancing TS modeling beyond short-term predictions.
81. ML-driven detection and reduction of ballast information in multi-modal datasets
Core Problem: Modern datasets often contain redundant or low-utility information ('ballast') that increases dimensionality, storage requirements, and computational cost without contributing meaningful analytical value.
Key Innovation: A generalized, multimodal framework for ballast detection and reduction across diverse data types, using various analytical techniques (entropy, mutual information, Lasso, SHAP, PCA, topic modelling, embedding analysis) and a novel 'Ballast Score' to prune significant portions of feature space with minimal or improved classification performance.
82. Characterizing the Predictive Impact of Modalities with Supervised Latent-Variable Modeling
Core Problem: The challenge of effectively utilizing multimodal learning approaches when data is incomplete, with modalities often missing, collected asynchronously, or available only for a subset of examples, and quantifying the individual predictive impact of these modalities.
Key Innovation: PRIMO, a supervised latent-variable imputation model that quantifies the predictive impact of any missing modality. It models missing data through a latent variable, enabling the use of incomplete training data and providing an instance-level variance-based metric to analyze the impact of missing modalities on predictions.
83. Forecasting Anomaly Precursors via Uncertainty-Aware Time-Series Ensembles
Core Problem: Most existing anomaly detection methods are reactive, identifying anomalies only after they occur, and lack the capability to provide proactive early warning signals in time-series data.
Key Innovation: Proposed FATE (Forecasting Anomalies with Time-series Ensembles), an unsupervised framework that detects Precursors-of-Anomaly (PoA) by quantifying predictive uncertainty from diverse time-series forecasting models, and introduced PTaPR, a new metric for holistic assessment of early warning capabilities.
84. StructCore: Structure-Aware Image-Level Scoring for Training-Free Unsupervised Anomaly Detection
Core Problem: Max pooling, the standard for converting anomaly score maps to image-level decisions in unsupervised anomaly detection (UAD), discards crucial information about the distribution and structure of anomaly evidence, leading to overlapping scores for normal and anomalous samples.
Key Innovation: Introduced StructCore, a training-free, structure-aware image-level scoring method that computes a low-dimensional structural descriptor from anomaly score maps and refines image-level scoring via diagonal Mahalanobis calibration, significantly improving AUROC scores on benchmark datasets by exploiting structural signatures.
85. AdvSynGNN: Structure-Adaptive Graph Neural Nets via Adversarial Synthesis and Self-Corrective Propagation
Core Problem: Graph neural networks suffer significant performance degradation when encountering structural noise or non-homophilous topologies, limiting their resilience and accuracy in node-level representation learning.
Key Innovation: AdvSynGNN, a framework combining multi-resolution structural synthesis, a transformer backbone with adaptive attention for heterophily, an adversarial propagation engine for connectivity alterations, and a label refinement scheme for robust node-level representation learning.
86. B$^3$-Seg: Camera-Free, Training-Free 3DGS Segmentation via Analytic EIG and Beta-Bernoulli Bayesian Updates
Core Problem: Existing interactive 3D Gaussian Splatting (3DGS) segmentation methods are impractical for low-latency use due to their reliance on predefined camera viewpoints, ground-truth labels, or costly retraining.
Key Innovation: B$^3$-Seg (Beta-Bernoulli Bayesian Segmentation for 3DGS), a fast, camera-free, and training-free method for open-vocabulary 3DGS segmentation that reformulates segmentation as sequential Beta-Bernoulli Bayesian updates and actively selects the next view via analytic Expected Information Gain (EIG), providing provable information efficiency.
87. TimeOmni-VL: Unified Models for Time Series Understanding and Generation
Core Problem: Time series modeling faces a sharp divide between numerical generation and semantic understanding, with existing models struggling to achieve both high-fidelity numerical output and comprehensive semantic interpretation.
Key Innovation: TimeOmni-VL, the first vision-centric framework that unifies time series understanding and generation through fidelity-preserving bidirectional mapping between time series and images (Bi-TSI) and understanding-guided generation, significantly improving both semantic understanding and numerical precision.
88. Distillation and Interpretability of Ensemble Forecasts of ENSO Phase using Entropic Learning
Core Problem: While ensemble forecasts of ENSO phase yield state-of-the-art skill, their interpretability is challenging, making it difficult to understand the underlying spatiotemporal dynamics and predictive information for long-range climate forecasting.
Key Innovation: Proposes a distillation framework to compress an ensemble of entropy-optimal Sparse Probabilistic Approximation (eSPA) models into a compact set of 'distilled' models, preserving forecast performance while enabling diagnostic tractability, and uses spatial importance maps and case studies to identify predictive information and trace ENSO evolution.
89. FR-GESTURE: An RGBD Dataset For Gesture-based Human-Robot Interaction In First Responder Operations
Core Problem: First Responders face increasing difficulties in disaster operations, and effective gesture-based human-robot interaction for Unmanned Ground Vehicle (UGV) control lacks specialized datasets.
Key Innovation: Introduces FR-GESTURE, the first RGBD dataset specifically for gesture-based UGV guidance by First Responders, comprising 12 refined commands and data collected from multiple viewpoints and distances to facilitate research in human-robot interaction for disaster response.
90. Multi-View 3D Reconstruction using Knowledge Distillation
Core Problem: Large Foundation Models for multi-view 3D reconstruction (e.g., Dust3r) require significant inference time and compute resources, limiting their application in tasks like Visual Localization.
Key Innovation: Proposes a knowledge distillation pipeline where a smaller student model (CNN or Vision Transformer based) is trained using 3D reconstructed points from a large teacher model (Dust3r), aiming to achieve replicable performance with reduced computational overhead for scene-specific 3D reconstruction.
91. Instance-Wise Adaptive Sampling for Dataset Construction in Approximating Inverse Problem Solutions
Core Problem: Learning-based approaches for inverse problem solutions often require large, general-purpose training datasets, leading to substantial data collection costs, especially when the prior has high intrinsic dimension or high accuracy is needed.
Key Innovation: Proposes an instance-wise adaptive sampling framework that dynamically allocates sampling effort based on the specific test instance, iteratively refining the training dataset to the geometry of the inverse map around each test instance, achieving significant gains in sample efficiency for approximating inverse problem solutions.
92. On the Sample Complexity of Learning for Blind Inverse Problems
Core Problem: Data-driven approaches for blind inverse problems, despite empirical success, often lack interpretability and rigorous theoretical guarantees, limiting their reliability in applied domains and hindering understanding of their performance characteristics.
Key Innovation: Provides a theoretical analysis of learning in blind inverse problems within the LMMSE framework, deriving closed-form optimal estimators, establishing equivalences with Tikhonov-regularized formulations, proving convergence results, and deriving finite-sample error bounds that quantify the impact of operator randomness.
93. Improved Object-Centric Diffusion Learning with Registers and Contrastive Alignment
Core Problem: Existing object-centric learning (OCL) methods using Slot Attention with pretrained diffusion models suffer from slot entanglement and weak alignment between object slots and image content, hindering robust object discovery and representation quality.
Key Innovation: Proposes Contrastive Object-centric Diffusion Alignment (CODA), which employs register slots to absorb residual attention and reduce interference, and applies a contrastive alignment loss to explicitly encourage slot-image correspondence, leading to improved object discovery and property prediction.
94. Learning PDE Solvers with Physics and Data: A Unifying View of Physics-Informed Neural Networks and Neural Operators
Core Problem: The field of physics-aware data-driven approaches for solving Partial Differential Equations (PDEs) lacks a unified perspective, hindering understanding of relationships, limitations, and appropriate roles of paradigms like Physics-Informed Neural Networks (PINNs) and Neural Operators (NOs).
Key Innovation: A unifying perspective is proposed for PINNs and NOs within a shared design space, organizing existing methods by what is learned, how physical structures are integrated, and how computational load is amortized, to facilitate reliable learning-based PDE solvers.
95. A Unifying Framework for Robust and Efficient Inference with Unstructured Data
Core Problem: Using neural networks to extract structured features from unstructured data introduces biases, researcher degrees of freedom, and reproducibility issues, making robust and efficient inference challenging, especially in fields like economics.
Key Innovation: Develops MAR-S (Missing At Random Structured Data), a semiparametric missing data framework that corrects for neural network prediction error using a validation sample, enabling unbiased, efficient, and robust inference with unstructured data for both descriptive and causal estimands, and addressing inference with aggregated/transformed predictions.
96. CFF-SCC: A Two-Stage Framework for Power Infrastructure Extraction From Airborne Transmission Corridor Point Clouds
Core Problem: Existing semantic segmentation methods for transmission corridors struggle to accurately identify minority-class components like insulators and jumper wires from airborne point clouds due to their complex geometric structures and significant scale variations, which is crucial for power line inspection and safe operation.
Key Innovation: Proposing CFF-SCC, a progressive two-stage framework for accurate segmentation and extraction of minor transmission components. It employs a cascaded feature fusion module for coarse segmentation and a spatial context calibration module with RANSAC-based geometric constraints for accurate extraction, demonstrating significant improvement in IoU scores for insulators and jumper wires across various terrains.
97. Low-Rank Gradient Guidance With Mutual-Guided Mamba for Hyperspectral Pansharpening
Core Problem: Existing deep learning pansharpening methods degrade without large datasets, and Transformer-based approaches incur quadratic computational complexity, limiting their fusion performance for hyperspectral images.
Key Innovation: Proposes MPSRNet-Diff, a two-stage framework that generates a prior HS image using a pretrained diffusion model (exploiting low-rank property and PCA for dimensionality reduction) and then uses a Mamba-based progressive superresolution network (MPSRNet) with mutual-guided Mamba blocks for efficient, global dependency modeling to enhance spatial details.
98. HC-XLSTM: A Dual-Branch Framework With Paired Blocks for Hyperspectral Image Classification
Core Problem: Hyperspectral image (HSI) classification struggles to efficiently capture both fine local spatial-spectral structures and long-range contextual dependencies, as CNNs have limited receptive fields and Transformers incur quadratic complexity.
Key Innovation: Proposes HCxLSTM, a dual-branch framework built on extended Long Short-Term Memory (xLSTM) with paired blocks for multidirectional scanning. It uses mLSTM for richer feature capacity, sLSTM for fine-grained gating, and a CrossmLSTM module for linear-time fusion, capturing both long-range context and local details.
99. Kolmogorov–Arnold Dynamic Gating Network With Dual-Domain Multiscale Feature Aggregation for Hyperspectral Stripe Noise Removal
Core Problem: Hyperspectral images are highly susceptible to stripe noise, which is difficult to separate from ground objects and background, and existing methods often fail to sufficiently integrate frequency-domain and spectral information while preserving original image structure.
Key Innovation: Proposes KAN-D2Net, a Kolmogorov–Arnold dynamic gating network with dual-domain multiscale feature aggregation. It models stripe noise in the frequency domain, integrates multiscale spatial features, uses a learnable graph structure for channel focus, and employs a KAN-based dynamic gating mechanism for adaptive nonlinear feature expression.
100. Super-Resolution of Remote Sensing Images via the Parallel Lattice Attention Network
Core Problem: Deep learning-based super-resolution models for remote sensing images struggle to balance capturing long-range dependencies with computational efficiency, as CNNs focus on local features and Transformers incur high computational costs.
Key Innovation: Proposes a parallel lattice attention network (PLAN) with a Parallel Lattice Attention Block (PLAB) that splits input feature maps into concurrent branches. These branches use LAUnits to simultaneously extract fine-grained details and global context, which are then fused via attention-based integration, maximizing computational efficiency and feature reuse.
101. Deep-Learning-Based Interactive Segmentation in Remote Sensing
Core Problem: Despite its potential for efficient human-computer interaction and high accuracy, interactive segmentation has seen limited application in remote sensing imagery, where it could greatly facilitate the analysis of complicated landscapes.
Key Innovation: Conducts a benchmark study of five state-of-the-art click-based interactive segmentation methods on high-resolution aerial imagery, introduces the cascade-forward refinement (CFR) approach to enhance results, and develops SegMap, a dedicated online tool for interactive segmentation of remote sensing data, fine-tuned for robust interactivity and adaptability.
102. AMPUNet: Hierarchical Attention Map Pyramid for Semantic Segmentation of Remote Sensing Images
Core Problem: Accurate semantic segmentation of high-resolution remote sensing imagery is challenging due to complex spatial patterns, extreme scale variations, and fine-grained details. Existing attention-based methods suffer from computational complexity, misalignment of multiscale representations, and loss of semantic information during upsampling.
Key Innovation: Introduces AMPUNet, a novel framework that constructs a hierarchical, coarse-to-fine attention map pyramid. It features a hybrid sparse attention framework, a dimension correspondence module for multiscale attention map alignment, and an attention map merging module with a CAW strategy to propagate and refine attention maps across layers, improving detail preservation and global understanding.
103. The Met Office Unified Model Global Atmosphere 8.0 and JULES Global Land 9.0 configurations
Core Problem: The continuous need to improve the accuracy, stability, and physical representation within global atmosphere and land surface models for both weather forecasting and climate simulations.
Key Innovation: Development and evaluation of the GA8GL9 configuration, integrating advancements in convection, riming, and land surface processes, leading to reduced errors, improved spatial structure in NWP, and a more accurate mean climate compared to previous versions.
104. A new sub-chunking strategy for fast netCDF-4 access in local, remote and cloud infrastructures, chunkindex V1.1.0
Core Problem: Inefficient access to large netCDF-4 datasets, especially for sub-parts (e.g., time series) and compressed chunks, across various storage infrastructures and network conditions, leading to unnecessary data transfer and processing overhead.
Key Innovation: Introduction of a new sub-chunking strategy and a library (chunk-indexing) that allows efficient retrieval of sub-parts from compressed netCDF-4 chunks without decompressing the entire chunk, demonstrating performance comparable to Zarr on S3 and suggesting it can avoid costly data reformatting.
105. Toward universal steering and monitoring of AI models
Core Problem: A lack of understanding regarding how modern AI models encode knowledge, which hinders improvements in model capabilities and the implementation of effective safeguards.
Key Innovation: Introduction of a robust and scalable method for extracting linear representations of concepts from various large-scale AI systems, enabling effective monitoring and steering of model outputs to improve AI performance and safety.
106. Brazil endangers global climate and health
Core Problem: Brazil's legislative actions prioritize financial gain over environmental conservation, violating constitutional and international environmental protection mandates, leading to the destruction of critical ecosystems.
Key Innovation: The paper advocates for Brazil's Supreme Federal Court to reaffirm constitutional environmental protection, preventing the reinterpretation of laws that legitimize ecosystem destruction and indirectly contribute to climate change impacts.
107. Estimation of fragment size distribution using existing models and modification for underground blasting in diverse geological condition
Core Problem: Accurate prediction of pre-blast fragment size distribution is critical for downstream operations and safety in underground blasting, but existing models like the modified Kuz-Ram model show significant deviations in diverse geological conditions.
Key Innovation: Evaluated and refined the modified Kuz-Ram model for tunnel blasting in different rock masses, developing new multiple non-linear regression equations and an empirical relationship correlating mean fragment size with Rock Mass Rating (RMR), Geological Strength Index (GSI), and joint spacing, significantly reducing prediction errors.
108. SWOT based intertidal topography mapping: A standalone method validated in a complex coastal region
Core Problem: Current SWOT satellite altimetry classification lacks a dedicated intertidal class, making it challenging to accurately separate intertidal pixels from water ones for intertidal topography mapping.
Key Innovation: Introduces a standalone methodology using SWOT level-2 HR pixel cloud data to identify intertidal pixels based on the probability distribution function of sea surface height anomalies, achieving RMSE below 30 cm and showing potential for detecting spatiotemporal morphological evolution.
109. Label-efficient outdoor 3D object detection via single click annotation from LiDAR point cloud
Core Problem: The prohibitively expensive cost of manual annotation for training 3D object detectors from LiDAR point clouds, especially for complete instance-level bounding box annotations.
Key Innovation: Presents SC3D, a label-efficient method requiring only a single coarse click per LiDAR frame, which uses a mixed pseudo-label generation module and a mix-supervised teacher-student network to achieve state-of-the-art performance with significantly reduced annotation cost (0.2%).
110. Integrating spatiotemporal similarity for robust gap-filling in continuous surface water mapping with uncertainty quantification
Core Problem: Continuous monitoring of surface water is severely hampered by data gaps in satellite imagery due to cloud cover, and existing gap-filling methods often lack reliance on multiple information sources and uncertainty assessment.
Key Innovation: The Spatio-Temporal Gap-Filling (STGF) framework integrates inundation frequency, spatial, and temporal similarity within a probabilistic structure, providing pixel-wise confidence scores and robust performance for reliable, high-frequency global surface water maps.
111. Mapping three decades of forest structural changes in Japan using Landsat time series and airborne LiDAR data
Core Problem: Accurately mapping nationwide forest structural changes over long periods is challenging due to the difficulty of large-scale LiDAR implementation and underutilization of rich temporal information from multiple change detection algorithms.
Key Innovation: Developed a scalable framework integrating large-scale airborne LiDAR and Landsat time series data with U-Net convolutional neural networks, leveraging multiple change detection algorithms (LandTrendr and CCDC) to map forest structural attributes across Japan annually for three decades with high accuracy.
112. Elucidating the hydrochemical and isotopic processes of surface and groundwater in response to river drying up and re-flowing in an alluvial-proluvial fan-plain transition zone
Core Problem: Limited attention to the mechanisms governing surface water-groundwater (SW-GW) interactions during river drying up and re-flowing processes, particularly in complex alluvial-proluvial fan-plain transition zones.
Key Innovation: Integration of water level, hydrochemistry, and δD-δ18O isotopes with Monte Carlo-based end-member mixing analysis (EMMA) to elucidate and quantify SW-GW transformations in response to river drying and re-flowing in an alluvial-proluvial fan, identifying distinct transformation zones and seasonal variations in SW-GW contributions, advancing understanding of hydrological processes and informing water resource management.
113. Study on mechanical properties of coal under microbial fluid and ScCO2 coupling action
Core Problem: The long-term mechanical stability of coal mass under the coupled action of supercritical carbon dioxide (ScCO2) and microbial fluid is inadequately characterized, which is crucial for coalbed methane recovery and carbon neutrality.
Key Innovation: Conducted triaxial compression experiments and acoustic emission (AE) monitoring, coupled with mineralogical and pore-fracture evolution tracking, to study mechanical properties and crack evolution of coal. Developed a novel damage-based constitutive model to assess damage and instability mechanisms of deep coal seams.
114. Multi-scale imaging and analysis of core-sized samples using successive image registration and machine learning for pore scale modeling
Core Problem: The challenge of generating large, fully resolved, high-resolution images of heterogeneous porous media for special core analysis (SCAL) and pore scale modeling due to field-of-view limitations.
Key Innovation: A novel hybrid multi-scale imaging and image super-resolution (SR) technique, employing successive image registration and an EDSR machine learning algorithm, to create highly resolved, large field-of-view composite images of core-sized porous media samples for accurate petrophysical property prediction.
115. Interpretable machine learning models for predicting seismic performance in the plastic hinge region of flexural DSCW components
Core Problem: Designing Double-Skin Composite Wall (DSCW) components is challenging due to complex parameters and calculations, making it difficult to directly predict their seismic performance.
Key Innovation: A data-driven framework using interpretable machine learning models (with SHAP analysis) is established to directly predict the bending-bearing capacity and deformability of DSCW plastic hinge regions from fundamental design parameters, providing design recommendations for optimal seismic performance.
116. Machine Learning Argument of Latitude Error Model for LEO Satellite Orbit and Covariance Correction
Core Problem: The compromised accuracy of LEO satellite orbit propagation and uncertainty quantification, particularly due to mismodeling of atmospheric drag, which limits the applicability of the Gaussian assumption for error distribution.
Key Innovation: A machine learning approach (using neural networks or Gaussian Processes) to correct error growth in the argument of latitude for LEO satellites, which improves orbit propagation accuracy and extends the utility of Vector Covariance Message (VCM) ephemerides to longer time horizons by focusing on dominant error growth dimensions.
117. SemCovNet: Towards Fair and Semantic Coverage-Aware Learning for Underrepresented Visual Concepts
Core Problem: Modern vision models exhibit Semantic Coverage Imbalance (SCI), a bias arising from long-tailed semantic representations where rare yet meaningful semantics are underrepresented, affecting how models learn and reason.
Key Innovation: SemCovNet, a novel model that explicitly learns to correct semantic coverage disparities by integrating a Semantic Descriptor Map (SDM), a Descriptor Attention Modulation (DAM) module, and a Descriptor-Visual Alignment (DVA) loss, enhancing model reliability and substantially reducing Coverage Disparity Index (CDI).
118. Xray-Visual Models: Scaling Vision models on Industry Scale Data
Core Problem: The challenge of developing a unified vision model architecture for large-scale image and video understanding that can be trained on industry-scale, noisy social media data while maintaining superior accuracy and computational efficiency.
Key Innovation: Xray-Visual, a unified vision model architecture trained on over 15 billion image-text pairs and 10 billion video-hashtag pairs using a three-stage pipeline (MAE, semi-supervised hashtag classification, CLIP-style contrastive learning) and an EViT backbone, achieving state-of-the-art performance across diverse benchmarks.
119. Beyond Message Passing: A Symbolic Alternative for Expressive and Interpretable Graph Learning
Core Problem: Graph Neural Networks (GNNs) suffer from fundamental limitations, including the 1-Weisfeiler-Lehman (1-WL) expressivity barrier and a lack of fine-grained interpretability, hindering their trustworthiness in high-stakes domains.
Key Innovation: SymGraph, a symbolic framework that replaces continuous message passing with discrete structural hashing and topological role-based aggregation, theoretically surpassing the 1-WL barrier, achieving superior expressiveness, 10x-100x speedups, and generating rules with superior semantic granularity.
120. Cross Pseudo Labeling For Weakly Supervised Video Anomaly Detection
Core Problem: Weakly supervised video anomaly detection needs to both localize anomalies and identify their categories using only video-level labels, requiring a balance between temporal precision and semantic discrimination.
Key Innovation: CPL-VAD, a dual-branch framework with cross pseudo labeling that combines a binary anomaly detection branch (for localization) and a category classification branch (for semantic recognition), exchanging pseudo labels to leverage complementary strengths.
121. Continual uncertainty learning
Core Problem: Robust control of mechanical systems with multiple, intertwined nonlinear dynamics and operating-condition uncertainties remains a fundamental challenge, leading to sub-optimal policies and poor learning efficiency in deep reinforcement learning (DRL).
Key Innovation: A curriculum-based continual learning framework that decomposes complex control problems into sequential tasks, acquiring strategies for each uncertainty without catastrophic forgetting, and integrates a model-based controller for accelerated, sample-efficient residual learning, demonstrated on active vibration control for automotive powertrains.
122. Polaffini: A feature-based approach for robust affine and polyaffine image registration
Core Problem: Medical image registration is dominated by intensity-based methods, while feature-based approaches, though theoretically desirable, have struggled with reliable feature extraction, despite recent deep learning advances.
Key Innovation: Polaffini, a robust and versatile feature-based framework for anatomically grounded image registration that leverages pre-trained deep learning segmentation models to extract reliable anatomical feature points (centroids), enabling efficient global and local affine matching and producing smooth polyaffine transformations for improved structural alignment and non-linear registration initialization.
123. EAGLE: Expert-Augmented Attention Guidance for Tuning-Free Industrial Anomaly Detection in Multimodal Large Language Models
Core Problem: Existing deep learning methods for industrial anomaly detection provide limited semantic explanations and Multimodal Large Language Models (MLLMs) often require costly fine-tuning or lack consistent accuracy improvements for this task.
Key Innovation: EAGLE, a tuning-free framework, integrates expert model outputs to guide MLLMs for accurate and interpretable industrial anomaly detection, improving performance across MLLMs without parameter updates and achieving results comparable to fine-tuning methods.
124. Variational Grey-Box Dynamics Matching
Core Problem: Bridging the gap between black-box deep generative models and interpretable but incomplete physics-based simulation models to learn dynamics from observational trajectories without ground-truth physics parameters or scalability/stability issues.
Key Innovation: A novel grey-box method that integrates incomplete physics models into generative models using a structured variational distribution within flow matching, employing two latent encodings (stochasticity/multi-modal velocity and physics parameters with a physics-informed prior), and an adaptation for second-order dynamics.
125. Variational inference via radial transport
Core Problem: Standard variational inference (VI) using Gaussian surrogates often fails to capture the correct radial profile of target distributions, leading to poor coverage.
Key Innovation: Proposes radVI, an algorithm that optimizes over radial profiles of distributions in VI, leveraging Wasserstein space and radial transport maps, providing theoretical convergence guarantees and acting as a cheap add-on to existing VI schemes.
126. The Anxiety of Influence: Bloom Filters in Transformer Attention Heads
Core Problem: Understanding the specific functions and mechanisms of individual attention heads within transformer models, particularly whether some perform membership testing.
Key Innovation: Identifies and characterizes specific transformer attention heads that function as 'membership testers' (similar to Bloom filters) across different LLMs, detailing their strategies, capacities, and generalization abilities, and distinguishing them from other head types.
127. Provably Explaining Neural Additive Models
Core Problem: Generating provably cardinally-minimal explanations (smallest subset of features sufficient for prediction) for neural networks is computationally infeasible for standard models.
Key Innovation: Develops a model-specific algorithm for Neural Additive Models (NAMs) that efficiently generates provably cardinally-minimal explanations using a logarithmic number of verification queries, outperforming existing methods for even relaxed explanation tasks.
128. Position: Evaluation of ECG Representations Must Be Fixed
Core Problem: Current benchmarking practices for ECG representation learning are too narrow (focused on arrhythmia) and use suboptimal evaluation metrics, leading to unreliable progress and misalignment with broader clinical objectives.
Key Innovation: Argues for fixing ECG representation learning evaluation by expanding downstream tasks (structural heart disease, patient forecasting), outlining best practices for multi-label imbalanced settings, and demonstrating that a random encoder can surprisingly match state-of-the-art pre-training.
129. LATA: Laplacian-Assisted Transductive Adaptation for Conformal Uncertainty in Medical VLMs
Core Problem: Medical Vision-Language Models (VLMs) lack calibrated uncertainty guarantees under domain shift, leading to large, inefficient, and class-imbalanced prediction sets when using split conformal prediction (SCP).
Key Innovation: Introduces LATA (Laplacian-Assisted Transductive Adaptation), a training- and label-free method that refines VLM zero-shot probabilities by smoothing them over an image-image k-NN graph, improving prediction set efficiency and class-wise balance while preserving SCP validity.
130. MASPO: Unifying Gradient Utilization, Probability Mass, and Signal Reliability for Robust and Sample-Efficient LLM Reasoning
Core Problem: Existing Reinforcement Learning with Verifiable Rewards (RLVR) algorithms for LLMs suffer from inefficient gradient utilization (hard clipping), insensitive probability mass (uniform ratio constraints), and asymmetric signal reliability (disparate credit assignment).
Key Innovation: Proposes MASPO (Mass-Adaptive Soft Policy Optimization), a unified framework that integrates differentiable soft Gaussian gating, a mass-adaptive limiter, and an asymmetric risk controller to address the identified challenges, leading to more robust and sample-efficient LLM reasoning.
131. A Theoretical Framework for Modular Learning of Robust Generative Models
Core Problem: Training large-scale generative models is resource-intensive and relies on heuristic dataset weighting, making it difficult to achieve robust performance across diverse data mixtures or train modularly.
Key Innovation: Presents a theoretical framework for modular generative modeling, proving the existence of a robust gate to combine pre-trained experts, showing modularity as a strong regularizer, and demonstrating potential to outperform monolithic models, with a scalable training algorithm.
132. GraphThinker: Reinforcing Video Reasoning with Event Graph Thinking
Core Problem: Multimodal Large Language Models (MLLMs) for video reasoning often suffer from hallucinations due to a lack of explicit causal structure modeling within and across video events.
Key Innovation: Proposes GraphThinker, a reinforcement finetuning-based method that constructs structural event-level scene graphs and enhances visual grounding to reduce hallucinations in video reasoning by incorporating these graphs as an intermediate thinking process and using a visual attention reward.
133. RetouchIQ: MLLM Agents for Instruction-Based Image Retouching with Generalist Reward
Core Problem: Training MLLM agents for instruction-based image retouching using reinforcement learning is challenging due to the lack of reliable, verifiable reward signals that can reflect the subjective nature of creative editing.
Key Innovation: Introduces RetouchIQ, a framework for instruction-based image retouching using MLLM agents guided by a novel generalist reward model (an RL fine-tuned MLLM) that evaluates edited results through generated metrics, providing high-quality, instruction-consistent gradients for RL.
134. Revisiting Weight Regularization for Low-Rank Continual Learning
Core Problem: In parameter-efficient continual learning (PECL) with low-rank adapters, weight regularization techniques (like EWC) remain underexplored as a means to mitigate task interference while keeping storage and inference costs constant.
Key Innovation: Proposes EWC-LoRA, a method that leverages a low-rank representation to estimate parameter importance for EWC, effectively mitigating task interference in PECL with pre-trained models, offering a computationally and memory-efficient solution superior to existing low-rank CL approaches.
135. Reverso: Efficient Time Series Foundation Models for Zero-shot Forecasting
Core Problem: Existing time series foundation models for zero-shot forecasting are performant but inefficient and expensive due to their large scale (hundreds of millions of parameters).
Key Innovation: Introduction of Reverso, a family of efficient time series foundation models that are orders of magnitude smaller, using small hybrid models (interleaving long convolution and linear RNN layers) to match the performance of larger transformer-based models, along with data augmentation and inference strategies.
136. FAMOSE: A ReAct Approach to Automated Feature Discovery
Core Problem: Feature engineering remains a critical yet challenging bottleneck in machine learning, particularly for tabular data, as identifying optimal features from an exponentially large feature space traditionally demands substantial domain expertise.
Key Innovation: Introduction of FAMOSE, a novel framework leveraging the ReAct paradigm to autonomously explore, generate, and refine features while integrating feature selection and evaluation tools within an agent architecture, achieving state-of-the-art performance for both regression and classification tasks.
137. Self-Evolving Multi-Agent Network for Industrial IoT Predictive Maintenance
Core Problem: Industrial IoT predictive maintenance requires real-time anomaly detection with interpretability and low computational demands, but traditional models are static and cannot adapt to evolving conditions, while LLM-based systems are too resource-intensive for edge deployment.
Key Innovation: SEMAS, a self-evolving hierarchical multi-agent system that distributes specialized agents across Edge, Fog, and Cloud computational tiers for real-time anomaly detection. It incorporates dynamic ensemble detection, PPO-driven policy optimization, LLM-based explainability, and federated knowledge aggregation, achieving superior performance, stability, and efficiency for industrial predictive maintenance.
138. Semi-Supervised Learning on Graphs using Graph Neural Networks
Core Problem: A rigorous theory explaining when and why Graph Neural Networks (GNNs) succeed in semi-supervised node regression remains lacking.
Key Innovation: The paper provides a sharp non-asymptotic risk bound for least-squares estimation over GNNs with linear graph convolutions and a deep ReLU readout, separating approximation, stochastic, and optimization errors, and characterizing performance when labels are scarce.
139. Anti-causal domain generalization: Leveraging unlabeled data
Core Problem: Existing domain generalization methods typically require labeled data from multiple training environments, limiting applicability when labels are scarce, especially under distribution shifts.
Key Innovation: The paper proposes methods for anti-causal domain generalization that leverage unlabeled data from multiple environments to estimate perturbation directions and penalize model sensitivity, achieving worst-case optimality guarantees under certain environment classes.
140. AutoNumerics: An Autonomous, PDE-Agnostic Multi-Agent Pipeline for Scientific Computing
Core Problem: Designing accurate numerical solvers for PDEs requires substantial mathematical expertise and manual tuning, while neural network-based approaches often incur high computational cost and limited interpretability.
Key Innovation: AutoNumerics, a multi-agent framework that autonomously designs, implements, debugs, and verifies transparent numerical solvers for general PDEs directly from natural language descriptions, using a coarse-to-fine execution strategy and a residual-based self-verification mechanism.
141. Multiple Object Detection and Tracking in Panoramic Videos for Cycling Safety Analysis
Core Problem: Conventional crash records are sparse for identifying cycling risk factors, and existing computer vision models struggle with distortions, small objects, and boundary continuity in panoramic videos, limiting their use for naturalistic cycling safety studies.
Key Innovation: Proposes a novel three-step framework for multiple object detection and tracking in panoramic videos, involving segmenting/projecting images for enhanced detection, modifying tracking models for boundary continuity and object categories, and validating with vehicle overtaking detection, demonstrating improved precision and reduced identification switches.
142. Point Linguist Model: Segment Any Object via Bridged Large 3D-Language Model
Core Problem: Existing 3D object segmentation methods using Large Language Models (LLMs) are hindered by representation misalignment between high-level semantic tokens of LLMs and dense geometric structures of 3D point clouds, limiting input and output accuracy.
Key Innovation: Proposes the Point Linguist Model (PLM) which bridges the representation gap using Object-centric Discriminative Representation (OcDR) for object-centric tokens and a Geometric Reactivation Decoder (GRD) for accurate segmentation by combining LLM-inferred geometry with dense features.
143. Inference-Time Search Using Side Information for Diffusion-Based Image Reconstruction
Core Problem: Existing diffusion-based image reconstruction methods for inverse problems typically overlook valuable side information, which could significantly improve reconstruction quality, especially in severely ill-posed settings.
Key Innovation: Proposes a novel inference-time search algorithm that guides the sampling process of diffusion models using diverse forms of side information (e.g., reference images, textual descriptions, anatomical MRI scans), acting as a plug-and-play framework that consistently improves reconstruction quality across various inverse problems without requiring any training.
144. LayerSync: Self-aligning Intermediate Layers
Core Problem: Improving the generation quality and training efficiency of diffusion models often relies on external guidance on intermediate representations, which can be complex or require additional resources.
Key Innovation: Proposes LayerSync, a self-sufficient, plug-and-play regularizer that improves diffusion model performance by self-aligning intermediate layers, using the model's own semantically rich representations to guide weaker ones, thereby enhancing generation quality and training efficiency across various modalities without external supervision or additional data.
145. Beyond Linear Surrogates: High-Fidelity Local Explanations for Black-Box Models
Core Problem: Existing local explanation methods for black-box machine learning models lack in generating high-fidelity explanations, failing to accurately capture the non-linear local behavior of complex models.
Key Innovation: Proposes a novel local model-agnostic explanation method using multivariate adaptive regression splines (MARS) to model non-linear local boundaries and N-ball sampling strategies to sample perturbed data directly from a desired distribution, leading to higher local surrogate fidelity.
146. Online Robust Reinforcement Learning with General Function Approximation
Core Problem: Most existing Distributionally Robust Reinforcement Learning (DR-RL) approaches rely on strong data availability assumptions (generative models, large offline datasets) and are restricted to tabular settings, limiting their applicability in online scenarios with general function approximation.
Key Innovation: Proposes a fully online DR-RL algorithm with general function approximation that learns robust policies solely through interaction, based on a dual-driven fitted robust Bellman procedure, and establishes sublinear regret guarantees characterized by a robust Bellman-Eluder dimension.
147. Accelerating Large-Scale Dataset Distillation via Exploration-Exploitation Optimization
Core Problem: Existing decoupling-based dataset distillation methods face an efficiency-accuracy trade-off: optimization-based methods achieve higher accuracy but demand intensive computation, while optimization-free methods are efficient but sacrifice accuracy.
Key Innovation: Proposes Exploration--Exploitation Distillation (E$^2$D), a method that minimizes redundant computation through full-image initialization and a two-phase optimization strategy (exploration for high-loss regions, exploitation for focused updates), achieving higher accuracy and significant speedup on large-scale benchmarks.
148. A Parameter-free Adaptive Resonance Theory-based Topological Clustering Algorithm Capable of Continual Learning
Core Problem: Traditional Adaptive Resonance Theory (ART)-based and topological clustering algorithms require manual specification of critical parameters (e.g., similarity threshold, edge deletion threshold), which significantly impacts performance and limits adaptability.
Key Innovation: Proposes an ART-based topological clustering algorithm that integrates parameter estimation methods for both the similarity threshold (using a determinantal point process-based criterion) and the edge deletion threshold (based on edge age), enabling superior clustering performance on synthetic and real-world datasets without requiring dataset-specific parameter specifications.
149. ReplaceMe: Network Simplification via Depth Pruning and Transformer Block Linearization
Core Problem: Large Transformer models, especially LLMs, are computationally expensive and resource-intensive, making their deployment and inference challenging, and existing pruning methods often require extensive retraining or fine-tuning.
Key Innovation: Introduces ReplaceMe, a generalized training-free depth pruning method that replaces transformer blocks with a linear operation estimated from a small calibration dataset. This approach achieves significant network simplification (up to 25% pruning) while maintaining high performance on LLMs with minimal computational overhead and no retraining.
150. Sufficient, Necessary and Complete Causal Explanations in Image Classification
Core Problem: Existing explanations for image classifiers often lack formal rigor, and logic-based explanations are not computable for general image classifiers due to strict assumptions.
Key Innovation: Demonstrating that causal explanations provide formal rigor and computability for image classifiers, introducing methods to identify sufficient, necessary, and -complete/1-complete components of an image's classification, and providing efficient, black-box algorithms for these explanations.
151. Three-dimensional Trajectory Reconstruction Method of Moving Targets Using Multi-channel Airborne WasSAR Based on L-Shaped Antenna
Core Problem: Existing airborne Wide-angle Staring Synthetic Aperture Radar with Ground Moving Target Indication (WasSAR-GMTI) methods exhibit low accuracy and poor robustness in the 3D trajectory reconstruction of moving targets.
Key Innovation: Proposes a 3D trajectory reconstruction method using a multichannel airborne WasSAR based on an L-shaped antenna. It uses an along-track subsystem with clutter suppression interferometry to extract motion parameters and a cross-track subsystem for cross-track interferometric processing, combining these to estimate 3D target positions and reconstruct trajectories.
152. MCDF-Net: Dynamic Adaptive Network Based on Modal Competition and Dual Encoder Feature Fusion for Remote Sensing Image Target Detection
Core Problem: Remote sensing image object detection in complex environments struggles with dynamically varying information quality from different modalities (infrared and RGB), requiring adaptive weight adjustment for each modality to achieve all-weather detection.
Key Innovation: Proposes MCDF-Net, a dynamic adaptive network based on modal competition and dual-encoder feature fusion. It features a hierarchical feature attention fusion module (HFAM) for global context and local detail fusion, and an information entropy-guided adaptive modal competition mechanism to filter high-confidence queries and dynamically determine the salient modality for targets, balancing modal contributions.
153. LSCNet: An Adaptive Cloud Detection Network via Local–Global Spatial Context
Core Problem: Accurate detection of thin and fragmented clouds in optical remote sensing images is challenging due to their low contrast and diverse morphology in complex scenes, degrading image quality for downstream applications.
Key Innovation: Proposes LSCNet, an adaptive cloud detection network that processes thin cloud features within global and local spatial contexts. It uses a Mamba-based multiscale fusion block with learnable fusion weights to integrate differential and complementary information across spatial and spectral dimensions, and a local gated Mamba block with a spatial gating mechanism to enhance detailed features and suppress background noise.
154. Soil drying with experimental warming depends on ecosystem type and warming method: First results of the Soil Warming to Depth Data Integration Effort (SWEDDIE)
Core Problem: Prior databases and syntheses of warming effects on ecosystems lacked comprehensive, depth-resolved, and temporally detailed data on soil temperature and moisture responses to warming, leading to site-specific biases and an incomplete understanding of context dependencies.
Key Innovation: Introduction of SWEDDIE, the first database characterizing whole soil profile warming responses across 26 diverse ecosystems, providing high temporal resolution and depth-resolved observations of soil temperature and moisture, revealing ecosystem- and method-dependent soil drying patterns.
155. Uncertainty sources in a large ensemble of hydrological projections: Regional Climate Models and Internal Variability matter
Core Problem: Quantifying and disentangling various sources of uncertainty (emission scenario, GCM, RCM, hydrological model, internal variability) in large ensembles of regional hydrological projections, especially for different flow indicators (low, mean, high flows).
Key Innovation: Uses the QUALYPSO method to quantify uncertainty sources in the Explore2 dataset for French river flow projections, showing the dominant sources for different flow types and highlighting the significant role of internal variability, which is often as large as or larger than climate change response uncertainty.
156. AI offers way to image and assess clinical cell samples
Core Problem: The need for rapid and efficient assessment of cellular characteristics in clinical diagnostic settings.
Key Innovation: Application of artificial intelligence to image and rapidly assess clinical cell samples, aiding diagnostic decisions.
157. AI succeeds in diagnosing rare diseases
Core Problem: The inherent difficulty and complexity in accurately diagnosing rare diseases.
Key Innovation: Development of an artificial intelligence system that successfully diagnoses rare diseases by integrating clinical data, genetic information, and literature searches, while providing underlying reasoning.
158. Who is using AI to code? Global diffusion and impact of generative AI
Core Problem: A comprehensive understanding of generative AI's impact on human innovation, equity, and productivity across different skill levels within software development was lacking.
Key Innovation: Developed a machine learning classifier to detect genAI-generated code, revealing that while entry-level developers used genAI most, experienced senior developers leveraged it to increase productivity and innovation, highlighting a growing skills gap.
159. Protect US fisheries with fair markets
Core Problem: A proposed legislation (HR 6150) seeks to redefine 'disaster' to include economic hardship from foreign competition, potentially misallocating disaster assistance funds intended for natural or technological disasters affecting fisheries.
Key Innovation: The paper argues for more effective protection of US fisheries through better regulations, increased import monitoring, and vigilant enforcement of antidumping laws, rather than co-opting natural disaster assistance funds for economic competition issues.
160. The science and practice of proportionality in AI risk evaluations
Core Problem: Establishing effective and appropriate methods for evaluating risks associated with Artificial Intelligence systems, particularly concerning proportionality.
Key Innovation: Exploration of the scientific principles and practical application of proportionality in the context of assessing and managing AI-related risks.
161. Urfa stone: Evaluating the durability of the oldest known building stone in history
Core Problem: The long-term performance and durability of Urfa stone, widely used in historic and contemporary structures, under varying environmental conditions remains unclear.
Key Innovation: Conducted an extensive experimental campaign on Urfa stone, characterizing its petrographic, geochemical, and physico-mechanical properties, revealing its susceptibility to salt crystallization and freeze-thaw cycling as primary drivers of mechanical weakening and structural failure.
162. Predicting groundwater storage from seasonal managed aquifer recharge: insights from machine learning and explainable AI techniques
Core Problem: Capturing transient responses of groundwater storage after Managed Aquifer Recharge (MAR) using Machine Learning (ML) models remains a challenge, yet these are essential for understanding water retention through dry periods.
Key Innovation: Uses ML (U-Net and XGBoost) to model transient MAR effects by decomposing groundwater storage time series into a MAR-response and a decay coefficient, accurately predicting these components (R2 > 0.82), and employs Explainable AI (SHAP values) to identify key factors controlling MAR effectiveness for planning and optimization.
163. Characterization of brittleness evolution of hot dry rock during cyclic thermal treatment
Core Problem: Geothermal energy extraction from hot dry rock is hindered by low permeability, and while liquid nitrogen fracturing is promising, the evolution of rock brittleness during cyclic thermal treatment is poorly understood.
Key Innovation: Introduced a new brittleness index derived from fracture surface area and energy release rate to quantify progressive changes in rock brittleness during cyclic thermal treatment, demonstrating its applicability for optimizing geothermal energy extraction by enhancing fracture network development.
164. Leveraging sUAS-Sentinel-2 synergy for cross-scale mapping of canopy cover and aboveground biomass across Mongolia and Kazakhstan
Core Problem: Conventional canopy cover (CC) and aboveground biomass (AGB) mapping studies lack extrapolation robustness and induce uncertainty when upscaling limited in situ samples to broad spatial extents for rangeland management.
Key Innovation: Proposes and tests a spatial cross-scale approach integrating sUAS imagery as a bridge to upscale in situ data to satellite-based estimates (Sentinel-2), achieving improved predictions for CC and AGB mapping across large grassland regions.
165. Thermal Conductivity and Electrical Resistivity Characterization of Unsaturated Soils: Modified SWCC Cell and Prediction Models
Core Problem: Lack of precise and simultaneous measurement capabilities and quantitative models to predict the coupled variation and interrelationship of thermal conductivity (TC) and electrical resistivity (ER) with changes in saturation across the full range of unsaturated soil conditions.
Key Innovation: Developed a modified SWCC cell (MSc) apparatus for simultaneous and precise measurement of SWCC, TC, and ER, and proposed a four-parameter sigmoid function model to quantitatively describe the nonlinear variation of TC and ER as functions of saturation, showing good agreement with experimental data across different soil types.