TerraMosaic Daily Digest: Feb 23, 2026
Daily Summary
This 2026-02-23 digest (251 selected papers from 1,413 deduplicated candidates) reflects a decisive shift toward mechanism-aware hazard intelligence rather than pattern recognition alone. In landslide science, top studies resolve process chains from trigger to propagation: impulse-wave loading in soil-rock banks, InSAR coherence/phase precursors for slope failure, and seismic feature sets that separate rock and ice-rock avalanche signals before full collapse. Flood research is similarly moving from static hazard maps to operational control, with data-assimilation flood reanalysis, atmospheric-river compound-event attribution, and game-theoretic coordination of transboundary and multi-reservoir systems. Earthquake papers emphasize rapid but physically defensible impact estimation, including rupture-geometry sensitivity, empirical fragility calibration from observed damage, and probabilistic containment performance under uncertainty. Across domains, AI contributions are strongest when tightly coupled to physics, uncertainty analysis, and deployment constraints; purely generic model papers remain lower-priority unless they deliver clear transfer to geohazard decision workflows.
Key Trends
The strongest technical signal is convergence: sensing, physics, and inference are being integrated into systems that are both predictive and operationally usable.
- From displacement curves to precursor diagnostics: New landslide and avalanche studies rely on multi-channel precursors (coherence, wrapped phase, seismic energy structure) to increase lead time and reduce false confidence from single-signal monitoring.
- Flood science is becoming control-oriented: Research focus is shifting from event description to decision architecture, including assimilation-driven reanalysis, coordinated reservoir operations, and explicit risk-benefit tradeoff surfaces.
- Probabilistic seismic assessment is gaining practical depth: Recent work combines monitoring evidence, rupture geometry, and fragility modelling to improve near-real-time impact estimates while preserving uncertainty transparency for critical infrastructure decisions.
- Engineering mitigation is increasingly material- and mechanism-specific: Advances in root reinforcement quantification, frozen-soil treatment, tunnel damage-zone estimation, and interface mechanics indicate a move toward design rules grounded in measurable physical response.
- Foundation-model methods are being filtered by transfer value: High-volume AI papers remain prominent, but the most relevant contributions are those that demonstrably transfer to geohazard pipelines through physics constraints, interpretability, and robust out-of-domain behavior.
Selected Papers
This digest features 251 selected papers from 1413 papers analyzed across multiple journals. Each paper has been evaluated for its relevance to landslide and broader geohazard research and includes links to the original publications.
1. Internal stress evolution and attenuation law of opposite soil-rock mixture banks subjected to landslide-generated impulse waves
Core Problem: The internal dynamic response and hydro-mechanical coupling mechanism of permeable, heterogeneous Soil-Rock Mixture (S-RM) banks under the impact of landslide-generated impulse waves (LGIWs) are poorly understood due to measurement decoupling difficulties, hindering accurate assessment of reservoir safety.
Key Innovation: Investigated the hydro-mechanical coupling and stress evolution within saturated S-RM banks using a novel decoupled experimental strategy. Identified an 'Input-Transfer-Output' wave energy mechanism, observed significant hysteresis and spectral filtering effects in internal effective stress, and characterized a 'three-point' stress distribution pattern where peak dynamic stress attenuates exponentially. Revealed a distinct deflection of the stress transmission path, providing a quantitative attenuation law and theoretical basis for assessing the dynamic stability of reservoir colluvial banks.
2. Beyond and beneath displacement time series: towards InSAR-based early warnings and deformation analysis of the Achoma landslide, Peru
Core Problem: Detecting precursors to slope destabilization with sufficient lead time and accuracy remains challenging, as traditional InSAR displacement time series often suffer from severe underestimations over landslides due to unwrapping issues.
Key Innovation: Demonstrates the potential of using alternative InSAR signal markers, specifically interferometric coherence and wrapped phase, to observe the deformation process and progressive failure of the Achoma landslide, detecting gravitational structures up to 5 years and critical shifts 3 months before failure, providing valuable lead time for early warning.
3. Mechanism and movement process of the '2.8' rock avalanche in Junlian County, Southwest China
Core Problem: Elucidating the disaster mechanisms and movement processes of the '2.8' rock avalanche in Junlian County, which caused significant loss of life, in a region prone to geohazards due to complex geological and environmental factors.
Key Innovation: Conducted field investigations, LiDAR mapping, and geological model analysis to identify key triggering factors (rainfall, earthquakes, anthropogenic activities, strong weathering, fractured rocks, loose deposits) and categorize three distinct types of rock avalanche movement processes: slope rock mass instability, cascading-scale amplification, and topography-constrained upheaval deposition.
4. Multi-parameter seismic metrics for detection and classification of rock and ice-rock avalanches
Core Problem: Early identification and precise monitoring of mass movements (rock and ice-rock avalanches) in high mountainous regions are challenging, especially for low-amplitude precursor and initiation signals that often overlap with ambient noise.
Key Innovation: Proposed a multi-parameter seismic metric (MSM) that quantifies instantaneous signal intensity, short-term energy, and cumulative energy trends, enabling efficient detection and classification of continuous seismic signals from rock and ice-rock avalanches, including main events and precursors, maintaining high sensitivity during low-amplitude stages and supporting early warning and risk assessment.
5. An analytical framework to assess static versus dynamic triggering of fault-slip rockbursts
Core Problem: The underlying mechanisms of fault-slip rockbursts, particularly the relative contributions of static and dynamic coseismic stress changes, are not fully understood, hindering accurate hazard assessment in deep underground excavations.
Key Innovation: A novel analytical framework that integrates linear elastic fracture mechanics, seismic source theory, and the Kirsch solution to diagnose and quantify static and dynamic coseismic stress perturbations, enabling the creation of rockburst hazard maps and improved susceptibility assessment.
6. TimeRadar: A Domain-Rotatable Foundation Model for Time Series Anomaly Detection
Core Problem: Current time series foundation models (TSFMs) are ineffective for unsupervised time series anomaly detection (TSAD) because they focus on learning prevalent, regular patterns in predefined domains, making it difficult to differentiate rare, irregular abnormal patterns that can resemble normal ones in the same domain.
Key Innovation: Introduces TimeRadar, an innovative TSFM built in a fractional time-frequency domain to support generalist TSAD across diverse unseen datasets. It uses Fractionally modulated Time-Frequency Reconstruction (FTFRecon) with a learnable fractional order to adaptively rotate time series, optimally differentiating normal and abnormal signals. A Contextual Deviation Learning (CDL) component further models local abnormalities, enabling effective TSAD.
7. Design and Implementation of a 25-Year Pseudo-Prospective Earthquake Forecasting Experiment in China (AoyuX)
Core Problem: Traditional evaluation of earthquake forecast models using log-likelihood scores of the full distribution P(n) can be biased by heavy tails and out-of-range observations, necessitating a more robust, tail-aware evaluation framework.
Key Innovation: Develops a tail-aware evaluation framework that estimates cell-wise P(n) using adaptive Gaussian kernel density estimation and performs a ~25-year pseudo-prospective earthquake forecast experiment (AoyuX) in China, demonstrating the sensitivity of model ranking to tail treatment and providing a robust estimate of the productivity exponent.
8. FlowCast: Advancing Precipitation Nowcasting with Conditional Flow Matching
Core Problem: Radar-based precipitation nowcasting faces challenges due to atmospheric dynamics uncertainty and the need for efficient modeling of high-dimensional data. While diffusion models produce reliable forecasts, their iterative sampling is computationally prohibitive for time-critical applications.
Key Innovation: FlowCast, the first end-to-end probabilistic model leveraging Conditional Flow Matching (CFM) for precipitation nowcasting, which learns a direct noise-to-data mapping in a compressed latent space, enabling rapid, high-fidelity sample generation, establishing new state-of-the-art in probabilistic performance and efficiency compared to diffusion models.
9. Laboratory experimental study on regular wave run-up on a hybrid vegetation-seawall
Core Problem: Attenuating wave run-up and protecting coastal zones from coastal flooding, specifically understanding the performance of hybrid vegetation-seawalls.
Key Innovation: Laboratory experimental study on wave run-up on a hybrid vegetation-seawall, investigating various influencing factors, and developing a prediction formula for eco-friendly coastal defenses.
10. Assimilation of SWOT Altimetry Data for Riverine Flood Reanalysis: From Synthetic to Real Data
Core Problem: Quantifying how satellite altimetry data, specifically from the SWOT mission, can improve hydraulic parameter estimation and river water level prediction in riverine flood reanalysis, especially when combined with in-situ gauges.
Key Innovation: Demonstrates that assimilating SWOT river node products, particularly in combination with in-situ water level measurements using ensemble Kalman filter strategies, significantly enhances the accuracy of riverine flood reanalysis and water level dynamics representation, reducing errors by an order of magnitude in OSSE and to below 17 cm in a real event.
11. Atmospheric Rivers as Triggers of Compound Flooding: quantifying Extreme Joint Events in Western North America Under Climate Change
Core Problem: Quantifying the extent to which Atmospheric Rivers (ARs) contribute to compound inland flooding (CIF) events, such as Rain on Snow (ROS) and Saturation Excess Flooding (SEF), in Western North America and understanding the influence of anthropogenic climate change relative to internal climate variability.
Key Innovation: Utilizes CanRCM4 Large Ensemble simulations to analyze the frequency and seasonality of AR-driven CIF events, demonstrating ARs as dominant drivers and highlighting the significant contribution of internal climate variability to future uncertainty, underscoring the need to integrate AR-related flooding risks into flood management strategies.
12. The 2020 Sivrice earthquake (MW6.8) and its seismotectonic linkage to the 2023 Kahramanmaraş earthquakes (MW7.8 and 7.6)
Core Problem: Understanding the seismotectonic linkage between the 2020 Sivrice earthquake and the catastrophic 2023 Kahramanmaraş earthquakes along the East Anatolian Fault Zone, specifically how the earlier event influenced the later ones.
Key Innovation: Employed field surveys, DInSAR analysis, moment tensor solutions, stress tensor analysis, and Coulomb stress change calculations to demonstrate that the 2020 Sivrice earthquake transferred stress to specific segments of the EAFZ and Southeastern Anatolian Thrust Zone, acting as a trigger for the 2023 Kahramanmaraş events and identifying remaining potential earthquake hazard sources.
13. A comparative analysis of tsunami flow through emergent and submerged coastal vegetation using numerical simulation
Core Problem: The need for effective mitigation strategies against strong tsunami-induced hydrodynamic forces threatening coastal regions, specifically understanding the comparative efficacy of emergent and submerged coastal vegetation in altering flow dynamics and dissipating energy.
Key Innovation: Employed Computational Fluid Dynamics (CFD) modeling with the volume of fluid technique to numerically compare the influence of submerged and emergent vegetation on tsunami flow behavior under varying Froude numbers. Found that emergent vegetation is superior in energy and velocity reduction but experiences higher fluid forces, while submerged vegetation is critical for near-bed flow regulation and provides better structural resilience, recommending a combined arrangement for optimal tsunami mitigation.
14. An improved GA-BP model based on self-organizing map mode unsupervised clustering ability promoting and its application in landslide susceptibility mapping
Core Problem: Traditional GA-BP models for regional landslide susceptibility prediction are disturbed by the spatial proximity of landslide and non-landslide samples, leading to significant prediction bias.
Key Innovation: Introducing the Self-Organizing Map (SOM) model's unsupervised clustering capabilities to mitigate bias in landslide susceptibility predictions when integrated with the GA-BP model (SOM-GA-BP), demonstrating superior predictive capability and reduced prediction bias.
15. Susceptibility assessment and sensitivity analysis of geological disasters in Chengdu–Chongqing urban agglomeration based on geographical detector model
Core Problem: Frequent geological disasters constrain socioeconomic development in the Chengdu–Chongqing urban agglomeration, necessitating accurate susceptibility evaluation and identification of sensitive factors for effective prevention and mitigation strategies.
Key Innovation: Development of a geological disaster susceptibility assessment model and sensitivity identification method based on the geographical detector model, screening key influencing factors (elevation, soil, geomorphology, vegetation) and their interactions to classify sensitivity levels and provide a scientific basis for precise prevention and control.
16. Precursory seismic patterns preceding major earthquakes—a case study from northern Himalayas
Core Problem: Forecasting major earthquakes (Mw ≥ 7.0) remains a significant challenge, requiring a better understanding of stress accumulation, dissipation, and seismicity along faults to identify reliable precursory patterns.
Key Innovation: Examining variations in seismic 'a' and 'b' values and energy release patterns to identify three stress evolution phases (Triggering, Accumulation, Release) preceding major earthquakes in the northern Himalayas, suggesting that continuous monitoring of these patterns can effectively forecast such events.
17. Evidence-based seismic performance assessment of non-structural components in fixed-base and base-isolated hospital buildings using monitoring data
Core Problem: The seismic vulnerability of non-structural hospital components (NSCs), critical for post-earthquake functionality, is often estimated using analytical models rather than quantified with observed earthquake data.
Key Innovation: This study provides a quantifiable, evidence-based assessment of real earthquake performance of fixed-base and base-isolated hospital buildings using multi-site strong-motion monitoring data, demonstrating that base-isolated hospitals consistently maintain floor accelerations and inter-story drift ratios below critical NSC limits.
18. Development of empirical fragility functions after the 2020 earthquakes in and around Türkiye
Core Problem: There is a need for comprehensive empirical fragility functions for predominant building typologies in Türkiye, based on observed damage data from recent major earthquakes, while accounting for uncertainties in ground motion models and local soil information.
Key Innovation: This paper constructs comprehensive empirical fragility functions for reinforced concrete and unreinforced masonry structures in Türkiye using statistical methods and extensive damage datasets from the 2020 Elazig and Izmir earthquakes, incorporating a logic tree approach for ground motion model uncertainties and providing comparisons with existing models.
19. Quantification of the mechanical reinforcement of roots of olive trees and their contribution to shallow slope stability
Core Problem: The mechanical effects of olive tree roots on shallow slope stability have not been thoroughly investigated, limiting understanding of their specific contribution to landslide prevention and land management strategies.
Key Innovation: Quantified the mechanical root reinforcement of olive trees using field measurements and a Root Bundle Model, integrating this into a probabilistic slope stability model to assess their impact on reducing shallow landslide susceptibility, showing that mature olive trees significantly enhance stability.
20. Determination of the in situ stress field in a fault zone of an underground gas storage reservoir: implications on aseismic fault slip and permeability enhancement
Core Problem: Underground gas storage (UGS) can reactivate pre-existing faults and cause leakage without detectable seismicity, and the mechanisms behind such aseismic fault slip and permeability enhancement, especially under depletion-altered stress paths, are not fully understood.
Key Innovation: An integrated approach combining field minifrac tests, lab rock-mechanics, 3D geomechanical modeling, reservoir simulations, and an analytical aseismic-slip model to diagnose cross-fault pressure communication in a UGS, demonstrating that depletion-altered stress paths promote aseismic sliding and transient permeability increases, leading to leakage even without seismicity.
21. Hydrologically driven coordination of transboundary floods
Core Problem: Transboundary river basins face complex flood management challenges due to propagation lags and multi-party interactions, leading to uncoordinated actions that exacerbate flood extremes.
Key Innovation: A hydrologically driven coordination framework integrating physical flood routing into an open loop differential game, which derives semi-analytical solutions for optimal strategies (e.g., upstream pre-release advance, staggered peak shaving) to reduce downstream flood peaks and improve basin welfare.
22. Understanding the tradeoff between the flood risk and the hydropower benefit from the integrated flood control operation of a reservoir group and a flood detention basin
Core Problem: The necessity to integrate reservoir group and flood detention basin operations for flood control, and to understand the inherent tradeoff between flood risk mitigation and hydropower benefits for enhanced decision-making.
Key Innovation: A framework to determine and understand the tradeoff between flood risk and hydropower benefit from the integrated flood control operation of a reservoir group and a flood detention basin, demonstrating how integrated operation enhances this exchange under extreme flood conditions and highlighting the critical impact of damage coefficients and safety flows.
23. Influence of rupture geometry in rapid earthquake impact assessment
Core Problem: Rapid earthquake impact assessments rely on accurate ground shaking estimates, but the influence of different rupture geometry definitions on these estimates and subsequent damage predictions is not fully understood.
Key Innovation: Explores different approaches to defining earthquake rupture geometry (from point-source to complex ruptures) and assesses their impact on estimated damaged buildings, finding that for low-to-moderate magnitude events, discrepancies can be significant but mitigated by using planar ruptures from basic seismogenic information or pre-computed databases.
24. Probabilistic analysis and fragility quantification of RC containment structures under seismic excitations using the probability density evolution method
Core Problem: Deterministic analyses of reinforced concrete containment structures (RCCS) under seismic events neglect significant uncertainty sources, leading to an inability to quantify the probability of failure and potentially underestimating damage.
Key Innovation: Develops a framework using the probability density evolution method (PDEM) combined with the equivalent extreme value event to perform probabilistic analysis and fragility quantification for RCCS under seismic excitations, proposing a new crack-width failure criterion and demonstrating significant differences from deterministic analysis.
25. A multi-scale deep learning framework for real-time post-earthquake damage estimation leveraging wavelet-driven GoogLeNet-FC
Core Problem: Existing post-earthquake damage assessment methods lack adaptability, are computationally intensive, or rely on limited ground motion metrics, hindering real-time, accurate, and spatially consistent predictions.
Key Innovation: Presents a rapid, multi-scale deep learning framework integrating wavelet transform with a GoogLeNet-FC network for real-time post-earthquake damage estimation, achieving high accuracy (96.2%), improved classification for transitional damage levels, and computational efficiency, validated by SHAP analysis and empirical benchmarks.
26. Experimental Study and Prediction of Frost Heave of Saline Silty Clay under Overburden Pressures
Core Problem: Previous research on frost heave and thaw settlement in saline silty clay often overlooks matric suction and the changes in unfrozen water within microscopic pores under overburden pressure, limiting the understanding of these processes in cold-region geotechnical engineering.
Key Innovation: Experimentally investigated the effects of overburden pressure on frost heave, thaw settlement, solution intake, matric suction, and pore water dynamics in saline silty clay using NMR and sensor measurements, and proposed a modified segregation potential model to predict frost heave rates by accounting for unfrozen water and salt content.
27. Recent Increasing Trend in October–November Caribbean Tropical Cyclone Activity
Core Problem: Understanding and quantifying recent trends in October–November Caribbean tropical cyclone activity and identifying the underlying climatic drivers influencing these trends.
Key Innovation: Identifying significant increasing trends in October–November Caribbean hurricanes, rapidly intensifying hurricanes, and landfalling hurricanes since 1979, and linking these to observed warming in the western Atlantic Warm Pool and anomalous cooling in the eastern Pacific, creating a more conducive environment.
28. NeXt2Former-CD: Efficient Remote Sensing Change Detection with Modern Vision Architectures
Core Problem: While State Space Models (SSMs) are gaining traction in remote sensing change detection, there's potential for modern convolutional and attention-based architectures to offer competitive or superior performance, especially in tolerating co-registration noise, spatial shifts, and semantic ambiguity in bi-temporal imagery.
Key Innovation: NeXt2Former-CD, an end-to-end framework integrating a Siamese ConvNeXt encoder, a deformable attention-based temporal fusion module, and a Mask2Former decoder, which achieves state-of-the-art results on multiple CD datasets, outperforming Mamba-based baselines in F1 and IoU while maintaining comparable inference latency.
29. InfScene-SR: Spatially Continuous Inference for Arbitrary-Size Image Super-Resolution
Core Problem: Standard diffusion-based image super-resolution models struggle to scale to arbitrary-sized images due to memory constraints and produce visible seams/inconsistent textures when processing large scenes via independent patches.
Key Innovation: InfScene-SR, a framework for spatially continuous super-resolution of large, arbitrary-sized scenes, which adapts diffusion models with a guided and variance-corrected fusion mechanism to seamlessly generate high-resolution imagery from remote sensing datasets, eliminating boundary artifacts and benefiting downstream tasks.
30. Make Some Noise: Unsupervised Remote Sensing Change Detection Using Latent Space Perturbations
Core Problem: Unsupervised change detection (UCD) in remote sensing struggles to generalize beyond a few change types in real-world scenarios because existing methods rely on predefined assumptions about change types introduced through handcrafted rules, external datasets, or auxiliary generative models.
Key Innovation: MaSoN (Make Some Noise), an end-to-end UCD framework, synthesizes diverse, data-driven changes directly in the latent feature space during training, dynamically estimated using target data statistics, enabling strong generalization across diverse change types and achieving state-of-the-art performance on multiple benchmarks, including SAR.
31. Adaptive Runge-Kutta Dynamics for Spatiotemporal Prediction
Core Problem: Existing deep learning approaches for spatiotemporal prediction struggle to effectively incorporate physical knowledge and accurately estimate the updating process of physical states, limiting their representative capacity.
Key Innovation: Introduces a physical-guided neural network that uses an adaptive second-order Runge-Kutta method with physical constraints and a frequency-enhanced Fourier module to precisely model and predict spatiotemporal dynamics, outperforming SOTA methods in various prediction tasks including weather forecasting.
32. Experiments and elastoplastic analyses on soil disturbance of soft clay subjected to cyclic loading
Core Problem: Soft clayey grounds can suffer significant damage from seismic events, but the mechanism of strength reduction due to cyclic loading and structural degradation is often overlooked.
Key Innovation: Conducting cyclic undrained triaxial tests on undisturbed soft clay to show significant shear strength reduction (up to 25%) due to structural degradation, and employing an elastoplastic constitutive model incorporating soil skeleton structure to successfully reproduce and explain these observations, providing a framework for evaluating post-earthquake performance and improving seismic design.
33. Development of an under-ice river discharge forecasting system in Delft-Flood Early Warning System (Delft-FEWS) for the Chaudière River based on a coupled hydrological-hydrodynamic modelling approach
Core Problem: Year-round river discharge estimation and forecasting in cold-climate regions is complicated by dynamic river-ice conditions, which alter channel hydraulics and invalidate open-water rating curves, with existing methods being site-specific and subjective.
Key Innovation: Develops and assesses a coupled hydrological-hydrodynamic modeling approach within the Delft-FEWS platform for under-ice river discharge forecasting, integrating meteorological products, a multi-model hydrological framework (HOOPLA), and 1D HEC-RAS river ice models to produce ensemble forecasts for discharge and water level.
34. Skills in sub-seasonal to seasonal terrestrial water storage forecasting: insights from the FEWS NET land data assimilation system
Core Problem: Accurate sub-seasonal to seasonal (S2S) terrestrial water storage (TWS) forecasting is critical for disaster response, but the factors influencing forecast skill, particularly model physics and initial conditions, need better evaluation.
Key Innovation: Evaluates S2S TWS forecasts from the FLDAS system (Noah-MP and CLSM models) over Africa, showing CLSM's superior performance due to reanalysis-based initial conditions that better capture interannual variability, and highlighting the critical role of independent observations (GRACE/FO) for improving TWS forecasts for disaster preparedness.
35. Ground surface deformation induced by grouting construction of shallow-buried shield tunnels in mudstone strata
Core Problem: Insufficient understanding and prediction of the quantitative influence of grout solidification and hardening on ground surface deformation during shield tunneling.
Key Innovation: Investigated how time-dependent grout solidification and hardening affect ground deformation using analytical solutions and numerical simulations, showing that explicit simulation yields larger settlement and grout hardening rate is decisive for control, with the shield tail being the critical location.
36. The essential function of low impact development facilities in mitigating urban flood disasters: approach for multi-source data fusion simulation
Core Problem: The exacerbation of urban flood disasters due to urbanization and the need to quantitatively evaluate the efficacy of Low Impact Development (LID) facilities in mitigating pluvial flooding.
Key Innovation: Developed a multi-source data fusion simulation approach using the Soil Conservation Service model, refined by incorporating LID stormwater retention capacities, to assess flood mitigation. Demonstrated that enhancing LID efficiency significantly reduces shallow inundation, expanding deployment reduces total inundation with diminishing returns, and optimizing layout improves regulation of both area and depth.
37. Regional W-phase solutions for past large-magnitude events in the Southwest Pacific
Core Problem: Rapid and stable earthquake characterization (magnitude, location, moment tensor) is paramount for Tsunami Early Warning (TEW), but traditional W-phase inversions can take up to 30 minutes, delaying critical information.
Key Innovation: Exploring regional W-phase inversion to reduce the time required to retrieve stable earthquake source information for TEW in NZ and the Southwest Pacific, demonstrating that reliable solutions can be obtained within 18 minutes, and magnitude estimation within 10 minutes, for large-magnitude earthquakes.
38. Shaking table test on the seismic response of subway station in the sand-fine mixture liquefiable site
Core Problem: Most studies on subway station seismic performance focus on clean sands, but naturally deposited sands are often mixed with fines, significantly influencing liquefaction resistance and requiring a better understanding of structural response in such liquefiable sites.
Key Innovation: Conducting shaking table tests to investigate the dynamic response and damage mechanism of subway stations in sand-fines mixture liquefiable sites under various seismic motions, revealing sensitivity of acceleration response to specific frequency contents and identifying center columns as weakest parts.
39. Thaw-driven permeability evolution of frozen soil with segregated ice: insights from low-field nuclear magnetic resonance and LBM simulations
Core Problem: There is a knowledge gap between mesoscopic features and macroscopic hydrological properties of unsaturated frozen soil during thawing, and a need to predict permeability evolution and understand thaw-induced soil settlement.
Key Innovation: Proposed a numerical thawing model based on the Lattice Boltzmann Method (LBM) to predict permeability evolution of thawing frozen soil with segregated ice, validated with NMR measurements. Demonstrated that thawing-induced permeability changes are related to surface temperature and porosity, and that the greatest liquid mass fraction change at the segregated ice layer is the primary driver of thaw-induced soil settlement.
40. Multi-performance parametric framework to enhance the design process and implementation of low-damage timber buildings
Core Problem: Optimizing the multi-performance design phase for innovative low-damage timber buildings (Pres-Lam technology) is critical to enhance their resilience against catastrophic events like earthquakes, while also considering energy efficiency and environmental footprint.
Key Innovation: A holistic integrated parametric framework was developed within Rhino-Grasshopper, utilizing Multi-Objective Optimization, to simultaneously consider seismic performance, energy efficiency, and environmental footprint in the design of Pres-Lam buildings, demonstrating its advantages in delivering sustainable and resilient structures.
41. A fourier-enhanced physics-informed Kolmogorov–Arnold network for multi-frequency seismic response analysis of structures
Core Problem: Conventional neural networks suffer from spectral bias, making it difficult to capture high-frequency features in structural responses under multi-frequency seismic excitations, and traditional numerical schemes are sensitive to noise or missing data.
Key Innovation: This study proposes a Fourier-enhanced Physics-Informed Kolmogorov–Arnold Network (FPIKAN) that integrates KAN's interpretability with Fourier-based input encoding and parameterized activation functions, achieving superior accuracy, stability, and robustness in multi-frequency seismic response analysis, even with noisy or low-frequency sampled data.
42. A new strategy for improving the anti-dispersal properties and mechanical performance of dispersed soil in seasonal frozen regions: Research on the application of soybean urease-induced carbonate precipitation (SICP)
Core Problem: Dispersive soils in seasonal frozen regions exhibit poor resistance to water erosion and deteriorate significantly under freeze-thaw cycles, posing severe geotechnical hazards, while conventional stabilization methods have ecological and long-term effectiveness issues.
Key Innovation: Proposed Soybean Urease-Induced Calcium Carbonate Precipitation (SICP) as an economical, environmentally friendly technique that significantly improves the anti-dispersion capacity and mechanical properties (e.g., 154.55% increase in UCS) of dispersive soils, and fundamentally enhances freeze-thaw resistance (271.34% higher strength after 30 cycles) by cementing soil particles and inhibiting pore formation.
43. Estimation of excavation-induced high-damage zone in super-large span tunnels with microseismic monitoring technology
Core Problem: Accurately estimating the excavation-induced high-damage zone is crucial for assessing rock mass stability in super-large span tunnels, but reliable and practical methods are needed.
Key Innovation: A new method to estimate the high-damage zone using microseismic monitoring technology, converting wave velocity reduction (obtained via double-difference imaging) into a damage state variable and spatially mapping it, providing a reliable tool for tunnel stability assessment.
44. Flood susceptibility prediction in the Indo-China Peninsula using 21 years of inundation occurrence data and machine learning
Core Problem: Accurately predicting flood susceptibility in large, data-scarce, and hydrologically complex regions like the Indo-China Peninsula, where traditional hydrological models face limitations and historical flood records are sparse.
Key Innovation: Developed a scalable, interpretable machine learning framework (XGBoost) for flood susceptibility prediction, integrating 21 years of dynamically extracted inundation occurrence data from MODIS imagery with twelve climate-land surface predictors, achieving high accuracy and providing insights into dominant predictors (NDVI, land use, temperature) through XAI techniques.
45. A novel hybrid deep learning model with dynamic parameterization for accurate flood simulation
Core Problem: Traditional process-based hydrological models are limited in capturing nonlinear runoff responses, while deep learning models lack physical consistency, leading to inaccurate flood simulation and forecasting crucial for disaster mitigation.
Key Innovation: A novel hybrid dynamic parameter network (HyDPNet) deep learning model that integrates a differential form of the Xinanjiang (XAJ) model into recurrent neural network units with parameters generated by auxiliary neural networks, combined with an LSTM post-processor, achieving superior flood simulation accuracy and physical interpretability.
46. Slowly Migrating Fracture Swarms in an Actively Serpentinizing Borehole
Core Problem: Understanding of fracture growth during serpentinization of peridotite, particularly in field settings, is limited to theoretical models and laboratory experiments.
Key Innovation: Provides the first field observations of slowly migrating tensile fracture swarms in actively serpentinizing peridotite boreholes, demonstrating that elevated pore pressure following rain events can trigger and sustain fracture propagation.
47. Multi‐Scale Electrical Conductivity Model of the Contiguous United States
Core Problem: Developing a comprehensive, continental-scale 3-D electrical conductivity model of the US asthenosphere-lithosphere structure from the USArray magnetotelluric dataset, which is computationally challenging.
Key Innovation: Presenting MECMUS, a novel multi-scale 3-D electrical conductivity model of the contiguous US derived from a single inversion of the USMTArray dataset, enabling accurate modeling of 3-D physics and unveiling coherent electrical signatures for various geological features and applications, including space weather hazard assessment.
48. Large Causal Models for Temporal Causal Discovery
Core Problem: Traditional causal discovery for temporal data follows a dataset-specific paradigm, limiting multi-dataset pretraining, struggling with large variable counts, and relying heavily on synthetic data, which hinders generalization.
Key Innovation: A principled framework for Large Causal Models (LCMs) that combines diverse synthetic generators with realistic time-series datasets, enabling scalable and effective temporal causal discovery with competitive or superior accuracy, particularly in out-of-distribution settings, and fast, single-pass inference.
49. Simple Model for the Estimation of the Bimodal SWCC of Highly Weathered Tropical Soils
Core Problem: The bimodal soil water characteristic curve (SWCC) of highly weathered tropical soils, crucial for understanding their hydraulic behavior, is complex and needs a simpler estimation model.
Key Innovation: Presenting a simple model based on nonlinear regression analyses to predict the drying SWCC of bimodal lateritic soils, using routine geotechnical data (porosity, liquid limit, relative aggregation, GSD parameters), achieving high accuracy (R2 values of 0.89-0.92).
50. Numerical modeling of wave-induced loads on coastal bridges using a Coupled Eulerian– Lagrangian analysis
Core Problem: Coastal bridges are vulnerable to wave loading during extreme weather events, but existing numerical studies often use scaled models and idealized wave crests, limiting the evaluation of cumulative loading and transitions across clearance and submergence states, thus failing to capture peak structural demand accurately.
Key Innovation: Leveraged a high-resolution 3D Coupled Eulerian–Lagrangian (CEL) model to quantify wave-induced forces on a slab-on-girder bridge under a simulated 100-year coastal storm event. The model, validated with experimental data, demonstrated that zero clearance produced the most severe vertical demand due to soffit-impact slamming, and that transient impulses, not captured by quasi-static assumptions, govern peak structural demand for low-clearance decks. This provides refined design inputs for assessing bridge vulnerability.
51. A Eulerian-Lagrangian method for local scour incorporating a novel sediment initiation
Core Problem: Existing numerical methods for local scour prediction often produce inaccurate scour hole morphologies and underestimate scour depth, leading to unreliable predictions for hydraulic structure safety.
Key Innovation: Developed a novel Eulerian-Lagrangian numerical simulation method (Local Scour Squares - LSS) that uses Deviatoric Strain Rate (DEV) for sediment initiation, demonstrating improved accuracy in simulating scour hole morphology and better computational efficiency compared to existing software.
52. A risk analysis model for Arctic escort operations based on STPA and Bayesian Networks
Core Problem: Assessing and understanding the complex risks associated with Arctic icebreaker escort operations, which involve multiple interacting environmental, vessel, human, and organizational factors, is challenging.
Key Innovation: Developed a risk analysis model integrating Systems-Theoretic Process Analysis (STPA) and Bayesian Networks to comprehensively assess risks (e.g., collision with ice, besetting in ice) in Arctic escort operations, identifying primary risk factors and optimal navigation periods.
53. Optimizing Long-Term Morphodynamic Predictions in Response to Dominant Forcing: A Case Study on the Belgian Continental Shelf.
Core Problem: Accurately predicting long-term morphodynamic evolution in mixed-energy coastal systems, which requires resolving key hydrodynamic processes and correctly representing forcing events, but determining the 'correct' events and optimizing model parameters remains challenging.
Key Innovation: Development of a hydro-morphodynamic modeling framework (TELEMAC suite) with a new metric (Ψ-factor) to quantify wave-tide interaction for MORFAC-based optimization, demonstrating that predictive skill is linked to the co-occurrence of energetic wave events with the spring-neap tidal cycle, and resolving model shortcomings by including wave nonlinearity and Stokes-drift effects.
54. SM-RDC: A Three-Step Downscaling Framework for Daily 1-km Seamless SMAP Product Generation Over the Tibetan Plateau
Core Problem: Soil moisture products from microwave remote sensing suffer from low resolution and incomplete coverage, limiting their utility in regional hydrology and precision agriculture.
Key Innovation: A three-step downscaling framework (SM-RDC) that reconstructs, downscales, and calibrates SMAP products from 9 km to 1 km resolution, using a temporal-spatial 3-D convolutional network for gap filling, an attention-CNN for downscaling with auxiliary variables, and a residual correction method, achieving high accuracy and seamless spatial coverage.
55. Author Correction: Global subsidence of river deltas
Core Problem: The original global river-delta subsidence study required formal correction to ensure technical accuracy of a widely cited hazard baseline.
Key Innovation: Publishes an author correction that refines the previously reported global subsidence assessment, improving reliability for downstream hazard interpretation and policy use.
56. Study of the hydraulic cavitation mechanism and aperture optimization of drilling in high-gas coal seams
Core Problem: Enhancing gas pre-extraction efficiency in low-permeability, high-gas coal seams and eliminating the danger of coal and gas outbursts.
Key Innovation: Proposed pressure relief zone and plastic zone area ratios as evaluation indices, used FLAC3D numerical simulation to investigate cavitation hole effectiveness, and determined an optimal radius range (0.4-0.5 m) for hydraulic cavitation boreholes to maximize pressure relief and permeability enhancement.
57. Mapping gully erosion susceptibility in Babai Basin, lumbini Province, Nepal, using machine learning algorithms
Core Problem: Gully erosion poses a significant threat to land stability, and accurately identifying prone areas is crucial for effective mitigation planning, which traditional methods may not achieve efficiently.
Key Innovation: Blending multi-source geospatial data with optimized machine learning models (specifically Random Forest) to effectively identify locations vulnerable to gully erosion, providing a practical susceptibility map for land use planning and erosion control.
58. Analysis of atmospheric variations using GNSS signal as atmospheric sensor (PWV-GNSS) in the extreme rainfall events in Rio Grande do Sul (Brazil) in 2024
Core Problem: The state of Rio Grande do Sul experienced an unprecedented climatic disaster with extreme accumulated precipitation leading to rapid and severe river level rises and widespread inundation, highlighting the need to understand and identify such events in near-real-time.
Key Innovation: Analyzing direct and indirect atmospheric measurements, including GNSS-derived Precipitable Water Vapor (PWV), to understand extreme rainfall events, demonstrating that GNSS-derived PWV serves as a near-real-time predictor of extreme rainfall with up to 60 minutes of lead time.
59. Asymmetric deformation characteristics and stress deviator distribution in roadways under principal stress reorientation
Core Problem: The underlying mechanisms of asymmetric deformation and failure in surrounding rock of mountain tunnels and deep mining roadways, especially under principal stress reorientation induced by fault-mining interaction, are poorly understood.
Key Innovation: Developed an analytical expression for the second invariant of the stress deviator (J2) considering principal stress reorientation, demonstrating that both the increase and reorientation of principal stresses govern J2 distribution and asymmetric failure. Proposed optimizing J2 state through adjustments in roadway size, geometry, and support systems to mitigate asymmetric failure.
60. Online triaxial–NMR monitoring of pore–fracture reorganization and creep response in coal under a mining-type stress path
Core Problem: Mining disturbances induce complex stress evolution in coal, leading to continuous reorganization of the pore-fracture structure (PFS) and time-dependent mechanical behavior (creep), which is challenging to monitor and understand in detail.
Key Innovation: An integrated online triaxial loading–low-field NMR system to monitor water-saturated coal specimens under a mining-type stress path, using T2-spectrum descriptors to characterize PFS evolution and creep response, revealing systematic PFS reorganization and stress-level dependent creep behavior.
61. HieraBoost-Q: interpretable karst discharge prediction from multi-site electrical conductivity with SHAP-based mechanism insights
Core Problem: Improving runoff prediction in complex and heterogeneous karst catchments, which is challenging due to the variability of conduit-fracture-matrix systems and recharge processes, and providing interpretable insights into the underlying hydrological mechanisms.
Key Innovation: Development of HieraBoost-Q, an interpretable and bias-aware hybrid framework combining multi-site electrical conductivity (EC), hierarchical XGBoost, and SHAP-based interpretation, which significantly improves discharge prediction accuracy (e.g., R2 from 0.907 to 0.987) and provides mechanistic insights into feature contributions and interactions, supporting flood warning and water management in karst systems.
62. Theoretical approach for predicting vertical deformation of underlying shield tunnels due to excavation unloading, incorporating soil consolidation
Core Problem: Accurately predicting the vertical deformation (heave) of underlying shield tunnels due to foundation pit excavation is challenging, especially when considering the significant impact of soil consolidation (reverse-consolidation driven by negative excess pore water pressure).
Key Innovation: Developed an innovative foundation flexibility matrix incorporating soil consolidation effects based on Biot’s theory, providing a time-domain solution for tunnel vertical deformation, validated against FEM and monitoring data, and offering a reliable means for predicting short-term and long-term uplift responses.
63. Weak-Form Evolutionary Kolmogorov-Arnold Networks for Solving Partial Differential Equations
Core Problem: Strong-form evolutionary neural networks for time-dependent Partial Differential Equations (PDEs) suffer from ill-conditioned linear systems due to pointwise residual discretization and unfavorable computational cost scaling.
Key Innovation: Proposed a weak-form evolutionary Kolmogorov-Arnold Network (KAN) that decouples the linear system size from the number of training samples for improved scalability, rigorously enforces boundary conditions, and provides a stable and accurate approach for solving PDEs, with potential relevance to scientific machine learning and engineering applications.
64. Adaptive Time Series Reasoning via Segment Selection
Core Problem: Existing time series reasoning approaches inefficiently encode entire time series into a fixed representation, regardless of relevance, leading to suboptimal performance, especially on rare event localization and multi-segment tasks.
Key Innovation: ARTIST, a controller-reasoner architecture that formulates time-series reasoning as a sequential decision problem, interleaving reasoning with adaptive temporal segment selection using reinforcement learning, improving average accuracy by 6.46 absolute percentage points.
65. Transformers for dynamical systems learn transfer operators in-context
Core Problem: Understanding the mechanism behind in-context learning in large-scale foundation models (Transformers) when adapting to and forecasting dynamical systems unseen during training.
Key Innovation: Discovery that attention-based Transformers apply a transfer-operator forecasting strategy in-context, lifting low-dimensional time series using delay embedding and identifying long-lived invariant sets to forecast unseen physical systems without retraining.
66. PhysConvex: Physics-Informed 3D Dynamic Convex Radiance Fields for Reconstruction and Simulation
Core Problem: Reconstructing and simulating dynamic 3D scenes with both visual realism and physical consistency remains a fundamental challenge, as existing neural representations (e.g., NeRFs, 3DGS) excel in appearance but struggle to capture complex material deformation and dynamics.
Key Innovation: Proposes PhysConvex, a Physics-informed 3D Dynamic Convex Radiance Field that unifies visual rendering and physical simulation. It represents deformable radiance fields using physically grounded convex primitives governed by continuum mechanics, introducing a boundary-driven dynamic convex representation and a reduced-order convex simulation using neural skinning eigenmodes.
67. Depth-Enhanced YOLO-SAM2 Detection for Reliable Ballast Insufficiency Identification
Core Problem: RGB-only object detection models (like YOLOv8) show limited reliability (low recall) in identifying ballast insufficiency in railway tracks, particularly in visually ambiguous scenarios, leading to potential safety issues.
Key Innovation: A depth-enhanced YOLO-SAM2 framework that integrates depth-based geometric analysis (with a sleeper-aligned depth-correction pipeline) and SAM2 segmentation with YOLOv8, substantially improving the recall and F1-score for detecting insufficient ballast in railway tracks.
68. L3DR: 3D-aware LiDAR Diffusion and Rectification
Core Problem: Range-view (RV) based LiDAR diffusion models achieve 2D photo-realism but neglect 3D geometry realism, often generating various RV artifacts such as depth bleeding and wavy surfaces, leading to inaccurate local geometry.
Key Innovation: Designs L3DR, a 3D-aware LiDAR Diffusion and Rectification framework that regresses and cancels RV artifacts in 3D space. It uses a 3D residual regression network to predict point-level offsets and a Welsch Loss to focus on local geometry, achieving state-of-the-art generation and superior geometry-realism across multiple benchmarks, applicable to different LiDAR diffusion models.
69. Learning from Complexity: Exploring Dynamic Sample Pruning of Spatio-Temporal Training
Core Problem: Training deep learning models for spatio-temporal forecasting is computationally intensive due to massive, often redundant datasets, with existing solutions overlooking data inefficiency.
Key Innovation: ST-Prune, a novel dynamic sample pruning technique that intelligently identifies and prunes less informative samples based on the model's real-time learning state, significantly accelerating training speed and improving efficiency for spatio-temporal forecasting while maintaining or enhancing model performance.
70. FUSAR-GPT : A Spatiotemporal Feature-Embedded and Two-Stage Decoupled Visual Language Model for SAR Imagery
Core Problem: Existing Visual Language Models (VLMs) perform poorly on Synthetic Aperture Radar (SAR) imagery due to its complex imaging mechanism, sensitivity to scattering features, and the scarcity of high-quality SAR text corpora, limiting their application in remote sensing.
Key Innovation: FUSAR-GPT is a VLM specifically for SAR, introducing a SAR Image-Text-AlphaEarth feature triplet dataset, embedding multi-source remote-sensing temporal features via 'spatiotemporal anchors' for dynamic compensation, and employing a two-stage SFT strategy to decouple knowledge injection and task execution, achieving state-of-the-art performance in SAR visual-language benchmarks.
71. Questions beyond Pixels: Integrating Commonsense Knowledge in Visual Question Generation for Remote Sensing
Core Problem: Automatically generated questions for remote sensing images are often simplistic and template-based, limiting their utility for real-world applications like question answering or visual dialogue systems, as they lack depth and commonsense knowledge.
Key Innovation: Proposes KRSVQG, a Knowledge-aware Remote Sensing Visual Question Generation model that integrates external knowledge triplets and uses image captioning as an intermediary representation to generate richer, more diverse, and knowledge-grounded questions for remote sensing images, even in low data regimes.
72. Knowledge-aware Visual Question Generation for Remote Sensing Images
Core Problem: Automatically generated questions for remote sensing images tend to be simplistic and template-based, hindering the real deployment of question answering or visual dialogue systems by lacking contextual understanding and external knowledge.
Key Innovation: Proposes KRSVQG, a knowledge-aware remote sensing visual question generation model that incorporates external knowledge related to image content and leverages image captioning as an intermediary representation to enhance the image grounding and quality of generated questions.
73. Spectral bias in physics-informed and operator learning: Analysis and mitigation guidelines
Core Problem: Physics-informed and operator learning methods (PINNs, PIKANs, neural operators) suffer from spectral bias, where low-frequency components of solutions to PDEs are learned significantly faster than high-frequency modes, and the interaction of this bias with optimization dynamics and loss formulations is poorly understood.
Key Innovation: A systematic investigation demonstrating that spectral bias is fundamentally dynamical, not just representational, and can be substantially mitigated by using second-order optimization methods and spectral-aware loss formulations, improving the recovery of high-frequency modes in various PDE types, including earthquake dynamics.
74. LEVDA: Latent Ensemble Variational Data Assimilation via Differentiable Dynamics
Core Problem: Long-range geophysical forecasts are fundamentally limited by chaotic dynamics and numerical errors, and existing data assimilation methods are either computationally expensive (classical variational smoothers) or enforce weak trajectory-level constraints and assume fixed observation grids (latent filtering).
Key Innovation: Proposes Latent Ensemble Variational Data Assimilation (LEVDA), an ensemble-space variational smoother operating in the low-dimensional latent space of a pretrained differentiable neural dynamics surrogate. It performs 4DEnVar optimization to jointly assimilate states and unknown parameters without adjoint code, accommodating irregular spatiotemporal sampling, and achieving improved assimilation accuracy and computational efficiency in geophysical benchmarks.
75. SenTSR-Bench: Thinking with Injected Knowledge for Time-Series Reasoning
Core Problem: A persistent gap exists in time-series diagnostic reasoning: General Reasoning Large Language Models (GRLMs) lack domain-specific knowledge, while fine-tuned Time-Series LLMs (TSLMs) lack the capacity to generalize reasoning for complex questions.
Key Innovation: A hybrid knowledge-injection framework that injects TSLM-generated insights directly into GRLM's reasoning trace, achieving strong time-series reasoning with in-domain knowledge, further leveraging reinforcement learning with verifiable rewards (RLVR) for unsupervised knowledge elicitation.
76. Physics-informed Active Polarimetric 3D Imaging for Specular Surfaces
Core Problem: 3D imaging of complex specular surfaces remains challenging in real-world, dynamic scenarios, with existing techniques either requiring multi-shot acquisition (slow) or deteriorating performance on high spatial frequency/large curvature surfaces.
Key Innovation: A physics-informed deep learning framework for single-shot 3D imaging of complex specular surfaces, which combines polarization cues and structured illumination through a dual-encoder architecture with mutual feature modulation to accurately and robustly infer surface normals with fast inference.
77. Interpolation-Driven Machine Learning Approaches for Plume Shine Dose Estimation: A Comparison of XGBoost, Random Forest, and TabNet
Core Problem: Conventional photon-transport-based calculations for plume shine dose estimation are computationally expensive, hindering rapid assessment critical for nuclear facility safety and radiological emergency response.
Key Innovation: An interpolation-assisted ML framework using XGBoost, Random Forest, and TabNet, augmented with shape-preserving interpolation, to achieve rapid and accurate plume shine dose estimation, with XGBoost showing highest accuracy and a web-based GUI for practical deployment.
78. Satellite-Based Detection of Looted Archaeological Sites Using Machine Learning
Core Problem: Monitoring thousands of remote archaeological sites for looting is operationally difficult, posing a severe risk to cultural heritage.
Key Innovation: Presents a scalable satellite-based pipeline using PlanetScope imagery and machine learning (ImageNet-pretrained CNNs with spatial masking) to detect looted archaeological sites, achieving high F1 scores and outperforming traditional ML methods.
79. NEXUS : A compact neural architecture for high-resolution spatiotemporal air quality forecasting in Delhi Nationa Capital Region
Core Problem: Urban air pollution in megacities like Delhi NCR poses critical public health challenges, requiring accurate and computationally efficient high-resolution spatiotemporal forecasting models for various pollutants (CO, NO, SO2).
Key Innovation: Introduces NEXUS, a compact neural architecture that integrates patch embedding, low-rank projections, and adaptive fusion mechanisms. NEXUS achieves superior predictive performance (R^2 > 0.91) for CO, NO, and SO2 with remarkably high computational efficiency (fewer parameters than state-of-the-art models), enabling real-time deployment for air quality monitoring.
80. Training Deep Stereo Matching Networks on Tree Branch Imagery: A Benchmark Study for Real-Time UAV Forestry Applications
Core Problem: Autonomous drone-based tree pruning requires accurate, real-time depth estimation from stereo cameras, but there's a lack of benchmarks for training deep stereo matching networks on real tree branch images for this specific application.
Key Innovation: Presents the first benchmark study training and testing ten deep stereo matching networks on the Canterbury Tree Branches dataset, identifying BANet-3D for best quality and AnyNet for near-real-time performance on UAV platforms, guiding resolution choices for forestry drone systems.
81. Unsupervised Anomaly Detection in NSL-KDD Using $\beta$-VAE: A Latent Space and Reconstruction Error Approach
Core Problem: The increasing integration of Operational Technology with Information Technology necessitates robust Intrusion Detection Systems, and there's a need to explore unsupervised anomaly detection methods for network traffic.
Key Innovation: Investigates an unsupervised anomaly detection approach using β-Variational Autoencoders on the NSL-KDD dataset, comparing the effectiveness of leveraging latent space structure (distances to training data projections) versus conventional reconstruction error for anomaly detection.
82. Brewing Stronger Features: Dual-Teacher Distillation for Multispectral Earth Observation
Core Problem: The diversity of Earth Observation (EO) sensors and modalities makes a single universal foundation model unrealistic, and existing EO pretraining methods (masked image modeling) emphasize local reconstruction but provide limited control over global semantic structure, hindering efficient knowledge transfer across modalities.
Key Innovation: A dual-teacher contrastive distillation framework for multispectral imagery is proposed, aligning the student's pretraining objective with modern optical vision foundation models by combining a multispectral teacher with an optical VFM teacher, enabling coherent cross-modal representation learning and achieving state-of-the-art results in both optical and multispectral settings for tasks like semantic segmentation, change detection, and classification.
83. RADE-Net: Robust Attention Network for Radar-Only Object Detection in Adverse Weather
Core Problem: Automotive perception systems struggle with robust object detection in adverse weather conditions due to optical sensor limitations, and existing Radar-based approaches often lose information by working with sparse point clouds or 2D projections.
Key Innovation: RADE-Net, a lightweight robust attention network tailored for 3D projections of 4D Range-Azimuth-Doppler-Elevation (RADE) tensors, which preserves rich Doppler and Elevation features while significantly reducing data size, achieving superior object detection performance in adverse weather.
84. Structured Bitmap-to-Mesh Triangulation for Geometry-Aware Discretization of Image-Derived Domains
Core Problem: Existing methods for discretizing image-derived boundaries for PDE simulation (e.g., CDT) can lead to global connectivity updates, non-determinism, and suboptimal mesh quality, hindering stable and efficient physically based simulations.
Key Innovation: A template-driven triangulation framework embeds raster/segmentation-derived boundaries into a regular triangular grid, retriangulating only intersected triangles, preserving the base mesh, and supporting parallel execution. It uses a finite symbolic lookup table for conflict-free retriangulation, resulting in closed meshes with bounded angles, fewer sliver elements, and improved geometric fidelity for PDE discretizations over image-derived domains.
85. Multivariate time-series forecasting of ASTRI-Horn monitoring data: A Normal Behavior Model
Core Problem: The need for a reliable model to forecast monitoring time-series data under normal operating conditions to enable early anomaly detection and predictive maintenance for complex systems.
Key Innovation: Develops a Normal Behavior Model (NBM) using a Multi-Layer Perceptron (MLP) for multivariate time-series forecasting of ASTRI-Horn telescope monitoring data, demonstrating consistent performance and effectiveness in providing reliable hour-scale predictions for early anomaly detection.
86. $R^2$-Mesh: Reinforcement Learning Powered Mesh Reconstruction via Geometry and Appearance Refinement
Core Problem: Existing mesh reconstruction methods from Neural Radiance Fields (NeRF) rely on limited training images, restricting supervision and making it difficult to fully constrain geometry and appearance, leading to suboptimal guidance for refinement.
Key Innovation: Proposes R2-Mesh, a reinforcement learning framework that combines NeRF-rendered pseudo-supervision with online viewpoint selection using a UCB-based strategy and geometry-aware reward, to jointly optimize SDF geometry and appearance for improved 3D mesh reconstruction.
87. SphOR: A Representation Learning Perspective on Open-set Recognition for Identifying Unknown Classes in Deep Learning Models
Core Problem: Deep Neural Networks (DNNs) struggle with Open-Set Recognition (OSR), often misclassifying unknown data as known classes, because existing methods don't explicitly optimize feature representations for identifying unknowns.
Key Innovation: Proposing SphOR, a supervised representation learning framework that explicitly shapes the feature space for OSR by enforcing discriminative class-specific features via orthogonal label embeddings, imposing a spherical constraint, and integrating Mixup/Label Smoothing, leading to state-of-the-art OSR performance.
88. Not All Pixels Are Equal: Confidence-Guided Attention for Feature Matching
Core Problem: Existing semi-dense feature matching methods treat all pixels equally during attention computations, introducing noise and redundancy from irrelevant regions.
Key Innovation: Proposes a confidence-guided attention mechanism that adaptively prunes attention weights based on precomputed matching confidence maps, introducing a confidence-guided bias and rescaling value features to attenuate uncertain regions, significantly enhancing feature matching performance.
89. FedSDAF: Leveraging Source Domain Awareness for Enhanced Federated Domain Generalization
Core Problem: Traditional Federated Domain Generalization (FedDG) methods overlook the unique knowledge embedded within source domains, leading to suboptimal generalization capabilities when adapting to unseen target domains, especially in strictly isolated federated learning environments.
Key Innovation: Proposes FedSDAF, the first systematic framework to enhance FedDG by leveraging source domain-aware features through a dual-adapter architecture (Domain-Aware Adapter and Domain-Invariant Adapter) and a Bidirectional Knowledge Distillation mechanism, significantly outperforming existing FedDG methods.
90. Performance Estimation in Binary Classification Using Calibrated Confidence
Core Problem: Traditional model performance monitoring requires ground truth labels, which are often unavailable or delayed, making it difficult to assess model performance (beyond accuracy) in deployed binary classification systems.
Key Innovation: Presents CBPE, a novel method that estimates any binary classification metric (accuracy, precision, recall, F1) without ground truth labels by treating confusion matrix elements as random variables and leveraging calibrated confidence scores to derive their distributions, providing estimates with strong theoretical guarantees and valid confidence intervals.
91. CodePDE: An Inference Framework for LLM-driven PDE Solver Generation
Core Problem: Solving Partial Differential Equations (PDEs) is complex, with traditional numerical solvers requiring expert knowledge and being computationally expensive, and neural-network-based solvers needing large datasets and lacking interpretability.
Key Innovation: Introduces CodePDE, the first inference framework for generating PDE solvers using large language models (LLMs) by framing PDE solving as a code generation task, demonstrating strong performance across a range of PDE problems with advanced inference-time algorithms and scaling strategies.
92. Select, then Balance: Exploring Exogenous Variable Modeling of Spatio-Temporal Forecasting
Core Problem: Existing spatio-temporal (ST) forecasting methods predominantly rely on a limited set of observed target variables, overlooking the potential of exogenous variables and facing challenges with their inconsistent effects and imbalance between historical and future data.
Key Innovation: ExoST, a general framework for exogenous variable modeling in ST forecasting that follows a 'select, then balance' paradigm, using a latent space gated expert module to dynamically select salient signals and a siamese dual-branch backbone with context-aware weighting for dynamic balance.
93. STCast: Adaptive Boundary Alignment for Global and Regional Weather Forecasting
Core Problem: Existing methods for regional weather forecasting from global atmosphere models are constrained by static and imprecise regional boundaries, leading to poor generalization ability and limiting the effectiveness of finer regional forecasts.
Key Innovation: Proposed STCast, a novel AI-driven framework for adaptive regional boundary optimization and dynamic monthly forecast allocation, employing a Spatial-Aligned Attention mechanism for boundary refinement and a Temporal Mixture-of-Experts module to capture temporal patterns, demonstrating consistent superiority in global/regional forecasting and extreme event prediction.
94. Efficient Generative AI Boosts Probabilistic Forecasting of Sudden Stratospheric Warmings
Core Problem: Accurate and efficient probabilistic forecasting of Sudden Stratospheric Warmings (SSWs) is challenging for traditional Numerical Weather Prediction (NWP) systems due to computational bottlenecks and limitations in physical representation, despite SSWs being key sources of subseasonal predictability and drivers of extreme winter weather.
Key Innovation: Develops FM-Cast, a Flow Matching-based generative AI model for efficient and skillful probabilistic forecasting of the spatiotemporal evolution of stratospheric circulation. FM-Cast forecasts SSW onset, intensity, and 3D morphology up to 15 days in advance, achieving comparable or superior skill to operational NWP systems with significantly reduced computational cost, and helps uncover distinct predictability regimes.
95. Rectifying Distribution Shift in Cascaded Precipitation Nowcasting
Core Problem: Existing cascaded deep learning models for precipitation nowcasting overlook the conflation of systematic distribution shift in deterministic predictions and local stochasticity, leading to inaccuracies in precipitation patterns and intensity, especially over longer lead times.
Key Innovation: RectiCast, a two-stage framework that explicitly decouples the rectification of mean-field shift from the generation of local stochasticity via a dual Flow Matching model, improving performance on radar datasets.
96. Safe and Near-Optimal Control with Online Dynamics Learning
Core Problem: The challenge of achieving both optimality and safety in control systems when system dynamics are unknown, particularly in real-world, non-episodic deployments.
Key Innovation: Introduces a framework for maximum safe dynamics learning that ensures provably safe operation throughout the learning process while continuously learning dynamics, and an algorithm to achieve near-optimal performance by learning dynamics only as needed, demonstrated in domains like drone navigation.
97. VIRTUE: Visual-Interactive Text-Image Universal Embedder
Core Problem: Existing multimodal embedding models lack visual-interactive capabilities to specify regions of interest (e.g., points, bounding boxes, masks) within images, limiting their applicability for localized grounding of user intent and entity-level information learning.
Key Innovation: Proposes VIRTUE, a Visual-InteRactive Text-Image Universal Embedder that integrates segmentation and vision-language models to process visual prompts for specific image regions, enabling more precise handling of complex scenarios and learning entity-level information, demonstrated with state-of-the-art performance on new benchmarks.
98. Temperature-dependent soil water retention and hydraulic conductivity model for biopolymer-treated soil
Core Problem: The influence of temperature on the hydraulic performance (water retention and hydraulic conductivity) of biopolymer-treated soils, which are emerging as eco-friendly soil improvement agents, is poorly understood.
Key Innovation: Developing a novel model to describe the temperature-dependent water retention and saturated hydraulic conductivity of biopolymer-treated soils, integrating effects on biopolymer water absorption, swelling, and hydrogel viscosity, and validating it with experiments. This offers a reliable tool for designing sustainable geotechnical systems under varying climatic conditions.
99. A Comprehensive Study on the Mechanical Performance of Geotextile-Wrapped Stone Columns under Triaxial Stress Conditions: Experiment, Numerical Simulation, and Analytical Model
Core Problem: A comprehensive understanding of the mechanical performance and failure mechanisms of geotextile-wrapped stone columns (GWSCs) under triaxial stress conditions is needed for their effective design and optimization.
Key Innovation: Conducting a comprehensive study integrating experiments, numerical simulation (coupled DEM-FDM with realistic gravel morphology), and an improved analytical model to understand GWSC performance, revealing a four-stage deformation mechanism, the role of force chains, and providing theoretical understanding and practical guidance for design.
100. The impact of low-level jets in the atmospheric boundary layer on the operation of large-scale floating offshore wind turbines
Core Problem: Low-level jets (LLJs) in the marine offshore atmospheric boundary layer cause irregular wind shear, significantly affecting the performance and safety of large-scale floating offshore wind turbines (OWTs) by increasing dynamic loads and structural response.
Key Innovation: A comprehensive analysis framework to evaluate LLJ effects on OWTs, modeling LLJ-induced wind profiles with polynomial functions and smoothed spectra, and using an Aero-Hydro-Servo-Structural (AHSS) coupled model to simulate OWT response, revealing significant increases in dynamic response and developing a rapid evaluation method for LLJ-induced effects.
101. Efficient Cloud Removal for Remote Sensing Data Transmission via Model Compression and Sparse Accelerator Design
Core Problem: Cloud coverage significantly degrades the quality and usability of remote sensing imagery, leading to unnecessary bandwidth consumption and power expenditure in satellite payloads.
Key Innovation: A high-efficiency hardware-software co-design framework for cloud removal, featuring a two-stage model compression strategy (reducing parameters by >98%) and a low-voltage sparse matrix accelerator (achieving 43.2% higher energy efficiency), validated across multisource datasets to reduce noninformative data transmission.
102. CSGANet: Lightweight Channel-Split Group Attention for High-Resolution Remote Sensing Change Detection
Core Problem: Remote sensing change detection (CD) methods face an accuracy–efficiency tradeoff, with heavy models being costly and lightweight ones eroding boundary fidelity, and lack a unified mechanism to fuse local, directional, and global context.
Key Innovation: Proposes CSGANet, a lightweight Siamese framework featuring a channel-split group attention (CSGA) module to selectively integrate complementary contexts and an Adaptive PathMixing Module (AdmPM) for globally guided, boundary-preserving downsampling, achieving competitive accuracy with reduced model complexity.
103. Spatial structures of emerging hot and dry compound events over Europe from 1950 to 2023
Core Problem: Understanding the historical changes in the probability of hot and dry compound events (CE) over Europe and North Africa, specifically identifying their spatial and temporal emergence from natural variability and attributing the drivers (changes in marginal distributions vs. dependence structure).
Key Innovation: Introduces the concept of 'Period of Emergence' (PoE) for in-depth signal analysis and uses bivariate copula models to decompose the contributions of changes in marginal distributions versus dependence structure to CE probability changes, revealing clear spatial patterns and the necessity of considering the dependence component for accurate emergence detection.
104. Effects of permeability, bacterial adsorption, and soil layering on entire MICP grouting process
Core Problem: Fully exploring the intricate hydraulic, biological, and chemical processes of MICP grouting for soil treatment through experiments alone is challenging, and the impacts of key soil properties (permeability, bacterial adsorption, layering) on the entire process need elucidation.
Key Innovation: Developed a multiphysics coupling model capturing fluid flow, transport, adsorption, reaction, and porosity/permeability variations to elucidate the entire MICP grouting process. Highlighted the critical role of bacterial distribution and soil properties (permeability, layering, bacterial attachment coefficient) in affecting raw material utilization efficiency and soil stabilization uniformity, providing fundamental understanding for designing effective MICP strategies.
105. The Applicability of Categorical Triple Collocation in Spatiotemporal Two-Dimensional Merging of Soil Freeze-thaw Products
Core Problem: Existing soil freeze/thaw (FT) products suffer from pronounced temporal and spatial inconsistencies and variable classification accuracy, leading to insufficient spatiotemporal stability, especially where ground observation data is sparse.
Key Innovation: Introduced a spatiotemporal two-dimensional (2D) merging approach with Categorical Triple Collocation (CTC)-derived dynamic weights to improve the spatiotemporal stability and accuracy of soil FT products, outperforming original and 1D merged products.
106. Identifying hotspots and impact factors of multi-type compound events over major global river basins
Core Problem: Identifying hotspots and understanding the impact factors of multi-type compound events (combinations of hazards like drought, heatwave, extreme precipitation) that amplify risks to globally interconnected socio-economic systems.
Key Innovation: Analyzed 12 types of compound events using global observations and reanalysis data across 520 major river basins to reveal hotspots (e.g., Eastern Asia, Mediterranean), frequencies, and seasonality, and determined key atmospheric circulation variability modes (ENSO, AO, NP) as significant impact factors, providing a basis for basin-scale climate risk management.
107. AquaCast: Urban water dynamics forecasting with precipitation-informed multi-input transformer
Core Problem: The challenge of accurately forecasting urban water dynamics, requiring models that can effectively integrate both endogenous and exogenous factors (like precipitation) and capture complex inter-variable and temporal dependencies.
Key Innovation: AquaCast, a multi-input, multi-output deep learning model that uses a transformer architecture to forecast urban water dynamics by capturing inter-variable and temporal dependencies across both endogenous (e.g., water height) and exogenous (e.g., precipitation) inputs, achieving significant improvements over baselines.
108. Evaluation of Atmospheric Models Over Mountainous Regions Using a Parsimonious Network Routing Model and Streamflow Observations: A Case Study of the Yarlung Zangbo River on the Tibetan Plateau
Core Problem: Evaluating kilometer-scale atmospheric models in data-sparse mountainous regions is challenging due to scarce in-situ observations, uncertain remote sensing products, and substantial uncertainty in hydrological models used to link precipitation to streamflow.
Key Innovation: Proposes a simple, low-uncertainty approach to evaluate atmospheric models by routing their runoff through a parsimonious network routing model calibrated against observed streamflow, providing complementary and more hydrologically relevant measures of atmospheric model performance in data-sparse mountainous regions.
109. An Empirical Model Combining Seismic Noise and Shear Stress to Predict Bedload Flux in a Gravel‐Bed Alluvial Channel
Core Problem: Bedload flux estimation, particularly at high temporal resolution, struggles to capture the intrinsic variability of transport in turbulent flow and the strong influence of local sediment size and morphological heterogeneity, despite existing hydraulics-based methods.
Key Innovation: Develops a new calibrated empirical equation combining seismic power spectral density (PSD) and excess shear stress to predict bedload flux at minute-scale resolution, which improves predictions by accounting for short- and medium-term flux variations, reduces scatter, and can reliably predict high-transport data even when calibrated on low-transport data.
110. Revisiting the Seasonal Trend Decomposition for Enhanced Time Series Forecasting
Core Problem: Time series forecasting presents significant challenges, and existing machine learning models can be enhanced by better addressing trend and seasonal components individually.
Key Innovation: Enhances machine learning models for multivariate time series forecasting by applying different strategies for trend and seasonal components, reducing error values and introducing computationally efficient dual-MLP models, with demonstrated improvements on a hydrological dataset.
111. Support Vector Data Description for Radar Target Detection
Core Problem: Classical radar detection techniques degrade in the presence of clutter (heavy-tailed distributions) and struggle when thermal noise combines with clutter, despite robust covariance estimators.
Key Innovation: Investigates and proposes two novel Support Vector Data Description (SVDD)-based detection algorithms (SVDD and Deep SVDD) as CFAR detectors, demonstrating their effectiveness on simulated radar data by avoiding direct noise covariance estimation.
112. Do Generative Metrics Predict YOLO Performance? An Evaluation Across Models, Augmentation Ratios, and Dataset Complexity
Core Problem: Reliably evaluating synthetic datasets for object detection before training is difficult, as standard generative metrics often do not predict downstream detection performance (mAP).
Key Innovation: Presents a controlled evaluation of synthetic augmentation for YOLOv11 across diverse detection regimes, benchmarking various generators and augmentation ratios. It identifies that metric-performance alignment is strongly regime-dependent and that many raw associations weaken after controlling for augmentation level.
113. Communication-Efficient Personalized Adaptation via Federated-Local Model Merging
Core Problem: Achieving communication-efficient personalized adaptation of large models in federated deployments, especially under task-level heterogeneity, where existing approaches lack theoretical justification for balancing general and personalized knowledge.
Key Innovation: Potara, a principled framework for federated personalization that constructs a personalized model by merging a federated model and a local model using closed-form optimal mixing weights derived from a variance trace upper bound, consistently improving personalization while reducing communication.
114. IRIS-SLAM: Unified Geo-Instance Representations for Robust Semantic Localization and Mapping
Core Problem: Existing dense geometric SLAM systems lack deep semantic understanding and robust loop closure, while semantic mapping approaches suffer from decoupled architectures and fragile data association.
Key Innovation: IRIS-SLAM, a novel RGB semantic SLAM system that leverages unified geometric-instance representations from an instance-extended foundation model to enable semantic-synergized association and instance-guided loop closure, significantly improving map consistency and wide-baseline loop closure reliability.
115. LaS-Comp: Zero-shot 3D Completion with Latent-Spatial Consistency
Core Problem: Existing 3D shape completion methods struggle with zero-shot, category-agnostic completion from diverse partial observations, often failing to effectively leverage the rich geometric priors of 3D foundation models.
Key Innovation: LaS-Comp, a training-free, two-stage framework, is introduced for zero-shot 3D completion, combining an explicit replacement stage to preserve partial observation geometry and an implicit refinement stage for seamless boundaries, outperforming state-of-the-art methods on a new comprehensive benchmark, Omni-Comp.
116. From Few-Shot to Zero-Shot: Towards Generalist Graph Anomaly Detection
Core Problem: Existing Graph Anomaly Detection (GAD) methods follow a 'one-model-for-one-dataset' paradigm, requiring dataset-specific training, leading to high computational/data costs, limited generalization, and challenges in privacy-sensitive scenarios.
Key Innovation: Proposes ARC, a few-shot generalist GAD method leveraging in-context learning for detecting anomalies on multiple unseen datasets with minimal retraining, and ARC_zero for zero-shot GAD using a pseudo-context mechanism, demonstrating strong generalization and efficiency.
117. Learning Multi-Modal Prototypes for Cross-Domain Few-Shot Object Detection
Core Problem: Open-vocabulary detectors for cross-domain few-shot object detection rely heavily on text prompts, missing domain-specific visual information crucial for precise localization of novel classes with limited examples.
Key Innovation: Proposes LMP (Learns Multi-modal Prototypes), a dual-branch detector that couples textual guidance with visual exemplars from the target domain, using a Visual Prototype Construction module and parallel text/visual-guided branches for improved precision and domain adaptation.
118. Robust Self-Supervised Cross-Modal Super-Resolution against Real-World Misaligned Observations
Core Problem: Performing cross-modal super-resolution on real-world data is challenging due to limited, misaligned low-resolution source and high-resolution guide image pairs, requiring robust self-supervised methods.
Key Innovation: Proposes RobSelf, a fully self-supervised model optimized online, featuring a misalignment-aware feature translator to align guide features and a content-aware reference filter for discriminative self-enhancement, achieving state-of-the-art SR performance on misaligned cross-modal data.
119. Open-Vocabulary Domain Generalization in Urban-Scene Segmentation
Core Problem: Existing Open-Vocabulary Semantic Segmentation (OV-SS) models struggle with robustness to domain shifts in unseen environments, particularly in urban-driving scenarios, due to distorted text-image correlations in pre-trained VLMs.
Key Innovation: Introduction of Open-Vocabulary Domain Generalization in Semantic Segmentation (OVDG-SS) as a new problem setting, along with S2-Corr, a state-space-driven text-image correlation refinement mechanism that mitigates domain-induced distortions, achieving superior cross-domain performance and efficiency.
120. Enhancing 3D LiDAR Segmentation by Shaping Dense and Accurate 2D Semantic Predictions
Core Problem: Semantic segmentation of 3D LiDAR point clouds, when reformulated as a 2D problem, suffers from sparse and inaccurate intermediate 2D semantic predictions due to the intrinsic sparsity of projected LiDAR and label maps, limiting final 3D accuracy in urban remote sensing.
Key Innovation: Develops MM2D3D, a multi-modal segmentation model leveraging camera images. Introduces cross-modal guided filtering to overcome label map sparsity and dynamic cross pseudo supervision to overcome LiDAR map sparsity, enabling dense and accurate 2D predictions that effectively enhance final 3D accuracy.
121. Marginalized Bundle Adjustment: Multi-View Camera Pose from Monocular Depth Estimates
Core Problem: Integrating dense Monocular Depth Estimation (MDE) with its significantly higher error variance into traditional Structure-from-Motion (SfM) pipelines is challenging, as SfM typically relies on sparse, triangulated point clouds.
Key Innovation: Introduces Marginalized Bundle Adjustment (MBA), a method that leverages the density of MDE depth maps to mitigate their error variance, enabling MDE to achieve state-of-the-art or competitive results in SfM and camera relocalization tasks across various scales.
122. OpenVO: Open-World Visual Odometry with Temporal Dynamics Awareness
Core Problem: Existing Visual Odometry (VO) methods struggle with varying observation frequencies and uncalibrated cameras, limiting their generalizability for real-world applications like extracting trajectories from diverse dashcam footage, as they overlook temporal dynamics information.
Key Innovation: Introduces OpenVO, a framework that explicitly encodes temporal dynamics within a two-frame pose regression and leverages 3D geometric priors from foundation models. This enables robust ego-motion estimation from monocular dashcam footage with varying observation rates and uncalibrated cameras, achieving significant performance and robustness improvements on autonomous driving benchmarks.
123. TeFlow: Enabling Multi-frame Supervision for Self-Supervised Feed-forward Scene Flow Estimation
Core Problem: Self-supervised feed-forward methods for scene flow estimation rely on unreliable two-frame point correspondences, which often break down under occlusions. Naive extensions to multi-frame supervision are ineffective due to inconsistent signals from abruptly varying correspondences.
Key Innovation: Introduces TeFlow, which enables multi-frame supervision by mining temporally consistent motion cues. It employs a temporal ensembling strategy to aggregate reliable supervisory signals from a candidate pool built across multiple frames, achieving state-of-the-art performance for self-supervised feed-forward methods with significant speedup on challenging datasets.
124. Training-Free Cross-Architecture Merging for Graph Neural Networks
Core Problem: Current model merging methods are constrained to homogeneous architectures, which is problematic for GNNs due to topology-dependent message passing and sensitivity to misalignment.
Key Innovation: Introduces H-GRAMA (Heterogeneous Graph Routing and Message Alignment), a training-free framework that lifts GNN merging from parameter space to operator space by formalizing Universal Message Passing Mixture (UMPM), enabling cross-architecture GNN merging without retraining and achieving inference speedups.
125. UP-Fuse: Uncertainty-guided LiDAR-Camera Fusion for 3D Panoptic Segmentation
Core Problem: LiDAR-camera fusion for 3D panoptic segmentation introduces a critical failure mode under adverse camera conditions (degradation, failure, calibration drift), compromising system reliability in safety-critical settings.
Key Innovation: UP-Fuse, a novel uncertainty-aware fusion framework that dynamically modulates cross-modal interaction using learned uncertainty maps, ensuring robustness under camera sensor degradation, calibration drift, and sensor failure, and employing a hybrid 2D-3D transformer for decoding.
126. TherA: Thermal-Aware Visual-Language Prompting for Controllable RGB-to-Thermal Infrared Translation
Core Problem: Large-scale data collection and annotation for thermal infrared (TIR) imaging is a major bottleneck, and existing RGB-to-TIR translation methods often overlook thermal physics, yielding implausible heat distributions.
Key Innovation: TherA, a controllable RGB-to-TIR translation framework that couples TherA-VLM with a latent-diffusion-based translator to produce diverse and thermally plausible images by encoding scene, object, material, and heat-emission context, allowing fine-grained control.
127. A Text-Guided Vision Model for Enhanced Recognition of Small Instances
Core Problem: Existing drone-based object detection models, like YOLO-World, need enhancement for precise recognition of small, specific targets, especially when guided by text prompts, and also require improved processing speed and efficiency.
Key Innovation: Introduces an improved YOLO-World model that replaces the C2f layer with a C3k2 layer for more precise local feature representation, optimizes parallel processing for speed and efficiency, and achieves superior accuracy and lightweight performance for small object detection in drone-based applications.
128. RAID: Retrieval-Augmented Anomaly Detection
Core Problem: Unsupervised Anomaly Detection (UAD) methods face challenges with noise due to intra-class variations, imperfect correspondences, and limited templates when matching test images to normal templates.
Key Innovation: Introduces RAID, a retrieval-augmented UAD framework that leverages retrieved normal samples from a hierarchical vector database to guide noise suppression in anomaly map generation, achieving state-of-the-art performance.
129. BayesFusion-SDF: Probabilistic Signed Distance Fusion with View Planning on CPU
Core Problem: Traditional 3D volumetric fusion techniques (e.g., TSDF) lack systematic uncertainty quantification and rely on heuristics, while neural implicit methods are GPU-intensive and less interpretable for decision-making.
Key Innovation: BayesFusion-SDF, a CPU-centric probabilistic signed distance fusion framework that models geometry as a sparse Gaussian random field, providing accurate 3D reconstruction with useful uncertainty estimates for active sensing and surface extraction.
130. One2Scene: Geometric Consistent Explorable 3D Scene Generation from a Single Image
Core Problem: Generating explorable 3D scenes from a single image is challenging, with existing methods producing severe geometric distortions and artifacts when viewpoints move far from the original perspective.
Key Innovation: Introduces One2Scene, a framework that decomposes the problem into panorama generation, explicit 3D geometric scaffold lifting via a generalizable Gaussian Splatting network (recasting as multi-view stereo), and novel view generation, enabling geometrically consistent and photorealistic explorable 3D scenes from a single image.
131. Bayesian Meta-Learning with Expert Feedback for Task-Shift Adaptation through Causal Embeddings
Core Problem: Meta-learning methods struggle with out-of-distribution target tasks, leading to negative transfer, and adaptation often relies on limited target-task data and noisy expert feedback.
Key Innovation: Proposes a causally-aware Bayesian meta-learning method that conditions task-specific priors on precomputed latent causal task embeddings, enabling transfer based on mechanistic similarity and mitigating negative transfer under task shift, demonstrated in simulations and clinical prediction.
132. A Computationally Efficient Multidimensional Vision Transformer
Core Problem: High computational and memory costs limit the practical deployment of Vision Transformers in a wide range of computer vision tasks.
Key Innovation: TCP-ViT, a novel tensor-based Vision Transformer framework built upon the Tensor Cosine Product, which exploits multilinear structures and orthogonality to enable efficient attention mechanisms and structured feature representations, achieving significant parameter reduction while maintaining competitive accuracy.
133. Learning Discriminative and Generalizable Anomaly Detector for Dynamic Graph with Limited Supervision
Core Problem: Dynamic graph anomaly detection (DGAD) is challenging due to the scarcity of labeled anomalies, leading to either ambiguous boundaries with unsupervised methods or overfitting/poor generalization with semi-supervised methods.
Key Innovation: An effective, generalizable, and model-agnostic framework for DGAD that learns a discriminative boundary from normal/unlabeled data while leveraging limited labeled anomalies. It employs residual representation encoding, a restriction loss, and a bi-boundary optimization strategy using a normalizing flow.
134. Flow3r: Factored Flow Prediction for Scalable Visual Geometry Learning
Core Problem: Current feed-forward 3D/4D reconstruction systems rely on expensive dense geometry and pose supervision, especially scarce for dynamic real-world scenes, limiting scalable training from unlabeled data.
Key Innovation: Introduces Flow3r, a framework that uses factored flow prediction as supervision to enable scalable visual geometry learning from unlabeled monocular videos, achieving state-of-the-art 3D/4D reconstruction results on static and dynamic scenes, with potential for deformation monitoring.
135. tttLRM: Test-Time Training for Long Context and Autoregressive 3D Reconstruction
Core Problem: Existing 3D reconstruction models struggle with long-context, autoregressive reconstruction and efficient processing of streaming observations, limiting their scalability and real-time application.
Key Innovation: Proposes tttLRM, a novel large 3D reconstruction model that uses a Test-Time Training (TTT) layer for long-context, autoregressive 3D reconstruction with linear computational complexity, enabling progressive refinement from streaming observations and achieving superior performance, with potential for deformation monitoring.
136. OVerSeeC: Open-Vocabulary Costmap Generation from Satellite Images and Natural Language
Core Problem: Generating global costmaps for long-range autonomous navigation directly from satellite imagery is challenging when mission requirements, terrain entities, and traversal rules are expressed in natural language and vary dynamically.
Key Innovation: OVerSeeC, a zero-shot modular framework that decomposes the problem into Interpret-Locate-Synthesize, using LLMs and open-vocabulary segmentation to generate executable costmap code from satellite imagery and natural language preferences, enabling flexible and mission-adaptive global planning.
137. Auto Quantum Machine Learning for Multisource Classification
Core Problem: Applying quantum computational methods to data-intensive scientific fields like remote sensing, particularly for complex data fusion and change detection tasks, requires efficient and effective QML approaches.
Key Innovation: Introduction of an automated QML (AQML) approach for multisource classification and data fusion, demonstrating improved accuracy over classical MLPs and manually designed QML models for change detection using multispectral remote sensing data.
138. WildOS: Open-Vocabulary Object Search in the Wild
Core Problem: Autonomous navigation in complex, unstructured outdoor environments requires robots to operate over long ranges without prior maps and with limited depth sensing, necessitating semantic reasoning about where to go and what is safe to traverse beyond purely geometric frontiers.
Key Innovation: WildOS is a unified system for long-range, open-vocabulary object search that combines safe geometric exploration with semantic visual reasoning. It builds a sparse navigation graph and utilizes ExploRFM, a foundation-model-based vision module, to score frontier nodes for traversability, visual frontiers, and object similarity, enabling robust and efficient navigation in diverse terrains.
139. Scale-PINN: Learning Efficient Physics-Informed Neural Networks Through Sequential Correction
Core Problem: Traditional Physics-Informed Neural Networks (PINNs) suffer from slow training and modest accuracy compared to modern numerical solvers, limiting their practical adoption in science and engineering.
Key Innovation: Introduces Scale-PINN, a learning strategy that integrates the iterative residual-correction principle from numerical solvers directly into the PINN loss formulation, achieving unprecedented convergence speed and superior accuracy across various PDE problems, including fluid dynamics and aerodynamics.
140. Rethinking Chronological Causal Discovery with Signal Processing
Core Problem: Causal discovery methods are sensitive to the mismatch between observation recording times and underlying event timings, specifically regarding sampling rate and window length, impacting their performance.
Key Innovation: Examines the sensitivity of classical and recent causal discovery methods to sampling rate and window length hyperparameters, demonstrating their impact and discussing how signal processing ideas can help understand these phenomena.
141. Geometry Distributions
Core Problem: Existing neural representations of 3D data struggle with challenges like handling thin structures and non-watertight geometries, limiting their flexibility and accuracy.
Key Innovation: Proposing a novel geometric data representation that models geometry as distributions using diffusion models and a new network architecture, which captures fine-grained geometric details without assumptions about surface genus or connectivity, demonstrating high geometric fidelity and potential for various 3D applications.
142. MEt3R: Measuring Multi-View Consistency in Generated Images
Core Problem: Traditional reconstruction metrics are unsuitable for measuring the quality of generated outputs from large-scale generative models for multi-view image generation, and there's a need for sampling-procedure-independent metrics, especially for multi-view consistency.
Key Innovation: Introducing MEt3R, a novel metric for multi-view consistency in generated images, which uses dense 3D reconstructions from image pairs (via DUSt3R) to warp and compare feature maps, providing a view-invariant similarity score for evaluating generative models.
143. Exploring Interpretability for Visual Prompt Tuning with Cross-layer Concepts
Core Problem: Current visual prompt tuning methods for adapting pre-trained visual foundation models lack interpretability, hindering AI reliability and knowledge discovery, as prompts are abstract embeddings.
Key Innovation: Introducing Interpretable Visual Prompt Tuning (IVPT), a framework that links visual prompts to human-understandable semantic concepts via cross-layer concept prototypes, enabling explanation of visual prompts at different network depths and semantic granularities.
144. Learning to Rank Critical Road Segments via Heterogeneous Graphs with Origin-Destination Flow Integration
Core Problem: Existing learning-to-rank methods for road networks fail to incorporate origin-destination (OD) flows and route information, limiting their ability to model long-range spatial dependencies for identifying critical road segments.
Key Innovation: Proposes HetGL2R, a heterogeneous graph learning framework that builds a tripartite graph unifying OD flows, routes, and network topology, and uses attribute-guided graphs with a heterogeneous joint random walk and Transformer encoding to learn embeddings for ranking road-segment importance.
145. SuperMAN: Interpretable and Expressive Networks over Temporally Sparse Heterogeneous Data
Core Problem: Real-world temporal data, such as sensor readings or medical records, often consists of multiple signal types recorded at irregular, asynchronous, and sparse intervals, making it challenging to effectively learn from and interpret such heterogeneous information.
Key Innovation: SuperMAN (Super Mixing Additive Networks), a novel and interpretable-by-design framework that learns directly from temporally sparse and heterogeneous signals by modeling them as sets of implicit graphs. SuperMAN provides diverse interpretability capabilities (node-level, graph-level, subset-level) and achieves state-of-the-art performance in high-stakes tasks, offering crucial insights into complex temporal data.
146. MOGS: Monocular Object-guided Gaussian Splatting in Large Scenes
Core Problem: State-of-the-art 3D Gaussian Splatting (3DGS) systems for large scenes primarily rely on costly LiDAR-based pipelines, leading to high memory/computation strain and limiting scalability, fleet deployment, and optimization speed.
Key Innovation: MOGS, a monocular 3DGS framework that replaces active LiDAR depth with object-anchored, metrized dense depth derived from sparse visual-inertial (VI) structure-from-motion (SfM) cues, achieving high-quality rendering with reduced training time and memory consumption.
147. RangeSAM: On the Potential of Visual Foundation Models for Range-View represented LiDAR segmentation
Core Problem: Traditional voxel/point-based LiDAR segmentation methods are computationally expensive and lack real-time efficiency, while range-view methods are underexplored, especially regarding the potential of Visual Foundation Models.
Key Innovation: Introduces RangeSAM, the first range-view framework adapting SAM2 for 3D LiDAR segmentation, incorporating architectural modifications to emphasize horizontal dependencies, customize configurations for spherical projections, and capture unique spatial patterns, achieving competitive performance with speed and scalability.
148. SAGE: Spatial-visual Adaptive Graph Exploration for Efficient Visual Place Recognition
Core Problem: Visual Place Recognition (VPR) requires robust retrieval of geotagged images despite large appearance, viewpoint, and environmental variation, and prior methods neglect the dynamic interplay between spatial context and visual similarity during training.
Key Innovation: Presented SAGE, a unified training pipeline that enhances granular spatial-visual discrimination by jointly improving local feature aggregation (Soft Probing module), organizing samples during training via an online geo-visual graph, and hard sample mining (greedy weighted clique expansion sampler), achieving state-of-the-art VPR performance.
149. Flower: A Flow-Matching Solver for Inverse Problems
Core Problem: Solving linear inverse problems to produce reconstructions consistent with observed measurements often requires complex methods or specific hyperparameters, limiting generalizability and ease of use.
Key Innovation: Introduced Flower, an iterative solver for linear inverse problems that leverages a pre-trained flow model for flow-consistent destination estimation, refinement onto a feasible set, and time-progression, approximating Bayesian posterior sampling and achieving state-of-the-art reconstruction quality with nearly identical hyperparameters across various problems.
150. Bayesian Network Structure Discovery Using Large Language Models
Core Problem: Traditional Bayesian network structure learning methods often require extensive observational data or rely on manual, error-prone incorporation of expert knowledge, making it difficult to analyze complex systems, especially in low- or no-data regimes.
Key Innovation: Introduces a unified framework for Bayesian network structure discovery that places LLMs at the center. PromptBN leverages LLM reasoning for data-free DAG generation with O(1) query complexity and global consistency. ReActBN refines this graph using statistical evidence and LLM-integrated ReAct-style reasoning for data-aware settings, outperforming prior baselines, particularly in low- or no-data regimes.
151. LocateAnything3D: Vision-Language 3D Detection with Chain-of-Sight
Core Problem: Current Vision-Language Models (VLMs) excel at 2D description but lack robust multi-object 3D detection capabilities, hindering their ability to perceive and act in the 3D world.
Key Innovation: LocateAnything3D, a VLM-native recipe that casts 3D detection as a next-token prediction problem using a 'Chain-of-Sight (CoS)' sequence, where the decoder emits 2D detections and then predicts 3D boxes under an easy-to-hard curriculum, preserving open-vocabulary and visual-prompting capabilities.
152. Foundation Model Priors Enhance Object Focus in Feature Space for Source-Free Object Detection
Core Problem: In Source-Free Object Detection (SFOD), domain shift degrades object-focused representations, leading to unreliable pseudo-labels and hindering adaptation to target domains.
Key Innovation: FALCON-SFOD, a framework that enhances object-focused adaptation by using Spatial Prior-Aware Regularization (SPAR) with foundation model priors (OV-SAM masks) to guide feature space, complemented by Imbalance-aware Noise Robust Pseudo-Labeling (IRPL).
153. Distributionally Robust Classification for Multi-source Unsupervised Domain Adaptation
Core Problem: Existing Unsupervised Domain Adaptation (UDA) approaches struggle when the target domain offers limited unlabeled data or spurious correlations dominate the source domain, leading to poor generalization.
Key Innovation: A novel distributionally robust learning framework that models uncertainty in both the covariate distribution and the conditional label distribution, applicable to multi-source and single-source UDA, with an efficient learning algorithm that integrates with existing UDA methods, showing improved performance especially when target data are extremely scarce.
154. Perception Characteristics Distance: Measuring Stability and Robustness of Perception System in Dynamic Conditions under a Certain Decision Rule
Core Problem: Current perception evaluation metrics for autonomous driving systems (ADS) do not adequately reflect the stochastic nature and uncertainty of AI perception algorithms, which is critical for safety in dynamic conditions.
Key Innovation: The Perception Characteristics Distance (PCD), a novel metric that incorporates model output uncertainty by representing the farthest distance at which an object can be reliably detected, along with the SensorRainFall dataset for empirical validation under various weather and illumination conditions.
155. Unleashing the Power of Discrete-Time State Representation: Ultrafast Target-based IMU-Camera Spatial-Temporal Calibration
Core Problem: Existing IMU-camera spatial-temporal calibration methods, relying on continuous-time state representation, suffer from high computational cost despite achieving precision.
Key Innovation: Proposes a novel and extremely efficient calibration method using discrete-time state representation, addressing its weaknesses in temporal calibration, leading to significant computational savings for visual-inertial platforms relevant for remote sensing.
156. BIMIM: Band-Independent Masked Image Modeling With Transformer for Multispectral Satellite Imagery
Core Problem: Existing Masked Image Modeling (MIM) frameworks, designed for natural images, struggle to adapt to the spectral-spatial characteristics of multispectral satellite imagery, with band-group embedding limiting flexibility and granularity of spectral feature learning.
Key Innovation: A novel Band-Independent Masked Image Modeling (BIMIM) with Transformer framework for multispectral satellite imagery, enabling finer band-specific spectral feature extraction and introducing spatially random masking at the single-band level for more efficient interband feature learning, achieving state-of-the-art performance in downstream tasks.
157. Contour Matching-Based Nondiffuse Reflection Noise Point Cloud Detection Method
Core Problem: Nondiffuse reflection noise in LiDAR point clouds negatively impacts the mapping accuracy of simultaneous localization and mapping (SLAM) technology.
Key Innovation: A contour-matching-based method for detecting nondiffuse reflection noise in point clouds, which projects point cloud clusters, extracts 2D contours using the alpha shape algorithm, constructs feature vectors, and identifies noise points via cosine similarity, achieving a correct removal rate of 91.44% and improved performance compared to similar approaches.
158. Self-Supervised Contrastive Learning for Hyperspectral Anomaly Detection With Block-Pseudolabel Masked Data Augmentation
Core Problem: Hyperspectral anomaly detection (HAD) suffers from a lack of labeled samples and often falls into partial anomaly representations, leading to high false alarm rates.
Key Innovation: Proposes an HAD method based on blind-block masked data augmentation and self-supervised contrastive learning, leveraging pseudolabels and a lightweight residual network to improve detection accuracy and background-anomaly separability while avoiding high computational costs.
159. Lightweight KAN Convolution Spectral–Spatial Network With Purification Window for Hyperspectral Anomaly Detection
Core Problem: Existing deep learning methods for hyperspectral anomaly detection (HAD) often reconstruct both background and anomalous pixels and neglect spatial information, making it difficult to separate anomalies from background.
Key Innovation: Introduces a purification window spectral–spatial self-supervised network that cleanses the dataset to reconstruct only background pixels, combined with a lightweight KAN convolution and depthwise separable convolution network and an improved guided filtering method to leverage spatial information and enhance HAD accuracy.
160. Sensitivity of hydrological machine learning prediction accuracy to information quantity and quality
Core Problem: While machine learning (ML) prediction accuracy in hydrology improves with more data, the specific mechanism and the role of information quantity and quality in this improvement are not well understood.
Key Innovation: Quantifies the connection between information amount (using Shannon's information theory) and ML prediction accuracy in hydrology, demonstrating that accuracy depends on both the quality and quantity of information in training data, and that combining weather data with calibrated theory-driven model outputs most efficiently improves accuracy.
161. Safe Anderson type-I least-squares reverse time migration based on a coefficient-optimized 25-point difference scheme
Core Problem: High computational costs of Least-squares reverse time migration (LSRTM) hinder its application for large-scale, high-resolution seismic imaging.
Key Innovation: Proposed a safe Anderson-type-I LSRTM, built upon an enhanced 25-point finite difference scheme, incorporating regularization and safety steps to improve stability, accelerate convergence, and achieve superior resolution and signal-to-noise ratio in seismic imaging.
162. Effects of biaxial confining stress on rock fragmentation and energy utilization in straight and empty hole blasting
Core Problem: Despite extensive research on confining stress effects in cut blasting, studies focusing on fragmentation characteristics and energy dissipation of deep confined blasting under biaxial confining stress remain scarce, hindering optimization for energy efficiency in mining.
Key Innovation: Integrated theoretical analysis, similarity model tests, and SPH-FEM simulations to investigate rock fragmentation size distribution (fitted by Swebrec function) and energy dissipation under varying biaxial confining stresses, demonstrating that increased confining stress increases fractal dimension and decreases fragmentation energy, providing insights for blasting efficiency in deep mining.
163. Investigation on Rock Fragmentation and Energy Dissipation Under Sandstone Blasting
Core Problem: There is a need for a comprehensive understanding of rock fragmentation characteristics and energy dissipation under sandstone blasting, particularly with varying powder factors, to optimize blasting efficiency in tunneling and mining.
Key Innovation: Experimentally and numerically investigated rock fragmentation (FSDs, fragment geometry) and energy dissipation under sandstone blasting with varying powder factors. Found that increased charging density leads to finer and more uniform fragmentation (well-characterized by extended Swebrec function) and identified an optimal specific charge for utilizing explosive energy in rock fragmentation.
164. Interlayer Slip Mechanisms in Bentonite and Their Influence on Mechanical and Swelling Behavior
Core Problem: Understanding bentonite's expansion and mechanical behavior under constant-volume conditions and varying moisture contents is crucial for its use as an engineered barrier, especially concerning excavation-induced fractures.
Key Innovation: This study proposes and validates a novel interlayer slip mechanism to explain bentonite's micro-deformation, demonstrating its influence on specific surface area, secondary swelling, and the transition from strain softening to strain hardening in direct shear tests, providing new insights for constitutive models.
165. A Nonlinear Load Transfer Model for Predicting the Torsional Response of Single Piles
Core Problem: Existing torsional load-transfer spring models for single piles under torsional loading exhibit limitations in adequately capturing soil nonlinearity.
Key Innovation: This paper proposes two sets of nonlinear torsional springs, particularly a modified hyperbolic model, for side and base resistances that accurately predict the torsional response of single piles, validated against full-scale case history data, without requiring correlations for Gmax values.
166. Shear Behavior of Paris Green Clay-Concrete Interface Under Constant and Cyclic Temperature Loadings
Core Problem: Daily and seasonal heating-cooling cycles in energy piles can trigger changes in the thermo-mechanical behavior of the soil-pile interface, significantly affecting design and performance.
Key Innovation: Direct shear tests using a specially designed apparatus revealed that constant cooling or heating temperatures increase adhesion and reduce the friction angle of the Paris green clay-concrete interface, while cyclic temperature loadings have a more limited influence, advancing understanding for energy pile design.
167. Bivariate-dependent remaining useful life prediction based on copulas and Tweedie exponential dispersion process
Core Problem: Existing Remaining Useful Life (RUL) prediction methods struggle with dependent failure modes, multiple degradation indicators, and various inherent uncertainties, limiting accuracy and lacking effective algorithms for simultaneous parameter updates.
Key Innovation: Proposed a novel generalized RUL prediction framework combining Tweedie exponential dispersion processes (for degradation modeling) and copula functions (for time-varying dependencies), along with an adaptive parameter estimation algorithm for online updating, achieving high accuracy in systems with complex multi-mode dependency and multiple uncertainties.
168. Dynamic water delay time estimation using dispatch data for improved inflow forecasting in cascade hydropower reservoirs
Core Problem: Accurate prediction of downstream inflows in cascade hydropower systems is fundamentally constrained by the dynamic and uncertain nature of water delay time, compounded by a prevalent lack of detailed hydraulic data.
Key Innovation: An operation-data-driven probabilistic framework that integrates an enhanced Dynamic Time Warping technique for high-resolution water delay time extraction, a Copula-Bayesian model for uncertainty quantification, and a probabilistic delay-matching scheme for forecasting, leading to superior robustness and accuracy in inflow prediction.
169. Hearing the forest for the trees: machine learning and topological acoustics for remote sensing with seismic noise
Core Problem: Monitoring remote forests is challenging due to limitations of satellite observations (weather, dense canopies, solar dependency).
Key Innovation: Demonstrates that passive seismic sensing can be used for autonomous ecosystem monitoring by capturing characteristic learnable signatures of trees within ambient seismic noise, achieving 86% classification accuracy for forest detection using machine learning and topological acoustics.
170. Depth from Defocus via Direct Optimization
Core Problem: Recovering depth from a collection of defocused images remains a computationally challenging optimization problem, despite a reasonable forward model for blur based on optical physics.
Key Innovation: Demonstrated that a global optimization approach using alternating minimization (convex optimization for all-in-focus image, parallel grid search for depth) can effectively solve the depth-from-defocus problem at higher resolutions than current deep learning methods.
171. JAEGER: Joint 3D Audio-Visual Grounding and Reasoning in Simulated Physical Environments
Core Problem: Current audio-visual large language models (AV-LLMs) are restricted to 2D perception, limiting reliable source localization and spatial reasoning in complex 3D environments.
Key Innovation: Introduces JAEGER, a framework extending AV-LLMs to 3D using RGB-D and multi-channel ambisonics, and proposes Neural IV for robust directional cues. It also provides SpatialSceneQA, a benchmark for training and evaluation.
172. Deep LoRA-Unfolding Networks for Image Restoration
Core Problem: Existing Deep Unfolding Networks (DUNs) for image restoration suffer from identical PMM architectures across stages (ignoring varying noise levels) and severe parameter redundancy/high memory consumption due to repetitive blocks.
Key Innovation: LoRun (Deep Low-rank Adaptation Unfolding Networks), which uses a single pretrained base denoiser shared across all stages with lightweight, stage-specific LoRA adapters injected into PMMs to dynamically modulate denoising behavior, achieving significant parameter reduction with comparable or better performance.
173. MiSCHiEF: A Benchmark in Minimal-Pairs of Safety and Culture for Holistic Evaluation of Fine-Grained Image-Caption Alignment
Core Problem: Fine-grained image-caption alignment in Vision-Language Models (VLMs) is crucial for socially critical contexts, such as identifying real-world risk scenarios, but current models struggle with subtle visual or linguistic clues, leading to potential misinterpretations.
Key Innovation: MiSCHiEF, a benchmark dataset with contrastive minimal-pair designs in safety (MiS) and culture (MiC) domains, is introduced to evaluate VLMs' ability to differentiate subtle visual and semantic distinctions, highlighting persistent modality misalignment challenges.
174. MIRROR: Multimodal Iterative Reasoning via Reflection on Visual Regions
Core Problem: Vision-Language Models (VLMs) often produce plausible but ungrounded answers, especially with ambiguous visual inputs, and their 'reflection' mechanisms may remain detached from image evidence, leading to hallucinations or logic errors.
Key Innovation: MIRROR, a framework for Multimodal Iterative Reasoning via Reflection On visual Regions, is introduced, embedding visual reflection as a closed-loop process (draft, critique, region-based verification, revision) repeated until the output is visually grounded, and a new dataset, ReflectV, is constructed for multi-turn supervision, demonstrating improved correctness and reduced visual hallucinations.
175. A high-resolution nationwide urban village mapping product for 342 Chinese cities based on foundation models
Core Problem: A consistent and reliable nationwide dataset for Urban Villages (UVs) in China is lacking due to their pronounced heterogeneity and diversity, hindering urban governance, renewal, and sustainable development.
Key Innovation: Presents GeoLink-UV, a high-resolution nationwide UV mapping product for 342 Chinese cities, derived from multisource geospatial data and a foundation model-driven framework, providing a reliable and systematically validated geospatial foundation for urban studies and informal settlement monitoring.
176. L2G-Net: Local to Global Spectral Graph Neural Networks via Cauchy Factorizations
Core Problem: Existing spectral Graph Neural Networks (GNNs) are either computationally expensive (full GFT) or limited in modeling long-range dependencies (local approximations), hindering their ability to capture global graph properties efficiently.
Key Innovation: Introduces L2G-Net, a new class of spectral GNNs that factorizes the Graph Fourier Transform (GFT) into operators on subgraphs combined via Cauchy matrices, enabling efficient processing of spectral representations of subgraphs to model both local and global dependencies with quadratic complexity.
177. SceneTok: A Compressed, Diffusable Token Space for 3D Scenes
Core Problem: Existing approaches for 3D scene representation and generation commonly use 3D data structures or view-aligned fields, which are less compressed and efficient for generation compared to a token-based approach.
Key Innovation: Presents SceneTok, a novel tokenizer for encoding view sets of scenes into a compressed and diffusable set of unstructured, permutation-invariant tokens, disentangled from the spatial grid. Achieves 1-3 orders of magnitude stronger compression than other representations while maintaining state-of-the-art reconstruction quality and enabling efficient scene generation.
178. Incremental Transformer Neural Processes
Core Problem: Existing Transformer Neural Processes (TNPs) are inefficient for sequential data streams, requiring recomputation of internal representations from scratch for every new observation, which is computationally expensive (quadratic time complexity).
Key Innovation: The Incremental Transformer Neural Process (incTNP), which leverages causal masking, Key-Value (KV) caching, and a data-efficient autoregressive training strategy to enable cheap, incremental updates with linear time complexity, matching or exceeding the predictive performance of standard TNPs for streaming inference.
179. Frame2Freq: Spectral Adapters for Fine-Grained Video Understanding
Core Problem: Existing time-domain adapters for adapting image-pretrained backbones to video often fail to capture dynamics across multiple time-scales, particularly medium-speed motion, which is crucial for fine-grained temporal analysis.
Key Innovation: Frame2Freq, a family of frequency-aware adapters that perform spectral encoding using Fast Fourier Transform (FFT) along time, learning frequency-band specific embeddings to adaptively highlight discriminative frequency ranges, thereby improving fine-grained action recognition in videos.
180. Direction-aware 3D Large Multimodal Models
Core Problem: Most existing 3D Large Multimodal Models (3D LMMs) and point cloud benchmarks lack corresponding ego poses for rich directional queries, making them inherently ill-posed for direction-aware spatial reasoning and question-answering.
Key Innovation: Redefines a paradigm for direction-aware 3D LMMs by identifying and supplementing ego poses into point cloud benchmarks and transforming data accordingly. It introduces PoseRecover (an automatic pose recovery pipeline) and PoseAlign (transforms point cloud data to align with identified ego poses), yielding consistent improvements across multiple 3D LMM backbones for tasks like ScanRefer and Scan2Cap.
181. Detecting labeling bias using influence functions
Core Problem: Labeling bias in datasets, arising from unequal label error rates or misrepresentation, makes fairness constraints ineffective and is challenging to detect.
Key Innovation: Investigates and demonstrates the use of influence functions to detect labeling bias by identifying wrongly labeled samples, showing promising results on MNIST and CheXpert datasets.
182. CaReFlow: Cyclic Adaptive Rectified Flow for Multimodal Fusion
Core Problem: The 'modality gap' significantly restricts the effectiveness of multimodal fusion, and existing methods often lack global distribution information or robust alignment.
Key Innovation: Proposes CaReFlow, extending rectified flow for multimodal distribution mapping using a 'one-to-many mapping' strategy, 'adaptive relaxed alignment', and 'cyclic rectified flow' to effectively reduce the modality gap and achieve competitive results in multimodal affective computing.
183. GS-CLIP: Zero-shot 3D Anomaly Detection by Geometry-Aware Prompt and Synergistic View Representation Learning
Core Problem: Current zero-shot 3D anomaly detection methods, which adapt CLIP by projecting 3D point clouds to 2D, inherently lose geometric details and provide an incomplete visual understanding, limiting their ability to detect diverse anomaly types.
Key Innovation: Introduces GS-CLIP, a framework that uses geometry-aware prompts embedded with 3D geometric priors (distilled by a Geometric Defect Distillation Module) and synergistic view representation learning (processing rendered and depth images in parallel with a Synergistic Refinement Module) to improve zero-shot 3D anomaly detection.
184. No Need For Real Anomaly: MLLM Empowered Zero-Shot Video Anomaly Detection
Core Problem: Existing video anomaly detection (VAD) methods underperform in open-world scenarios due to limited dataset diversity, rare occurrence of real anomalies, and inadequate understanding of context-dependent anomalous semantics.
Key Innovation: LAVIDA, an end-to-end zero-shot VAD framework that uses an Anomaly Exposure Sampler to generate pseudo-anomalies for training and integrates a Multimodal Large Language Model (MLLM) for enhanced semantic comprehension, along with a token compression approach for efficiency.
185. DefenseSplat: Enhancing the Robustness of 3D Gaussian Splatting via Frequency-Aware Filtering
Core Problem: 3D Gaussian Splatting (3DGS) is vulnerable to adversarial corruptions in input views, leading to degraded rendering quality, increased training/rendering time, and memory usage.
Key Innovation: Designs DefenseSplat, a frequency-aware defense strategy that reconstructs training views by filtering high-frequency noise while preserving low-frequency content, effectively suppressing adversarial artifacts and enhancing 3DGS robustness without clean ground-truth supervision.
186. Spiking Graph Predictive Coding for Reliable OOD Generalization
Core Problem: Existing Graph Neural Networks (GNNs) struggle with pervasive out-of-distribution (OOD) shifts in dynamic web environments, leading to unstable or overconfident predictions and lacking principled, interpretable uncertainty estimation.
Key Innovation: Introduces SpIking GrapH predicTive coding (SIGHT), an uncertainty-aware plug-in graph learning module that performs iterative, error-driven correction over spiking graph states to expose internal mismatch signals, thereby enhancing predictive accuracy, uncertainty estimation, and interpretability for reliable OOD generalization.
187. Redefining the Down-Sampling Scheme of U-Net for Precision Biomedical Image Segmentation
Core Problem: U-Net architectures for biomedical image segmentation often struggle to capture long-range information due to conventional down-sampling techniques that prioritize computational efficiency at the expense of information retention.
Key Innovation: Introduces 'Stair Pooling,' a simple yet effective strategy that moderates the pace of down-sampling by leveraging a sequence of concatenated small and narrow pooling operations in varied orientations, reducing dimensionality reduction from 1/4 to 1/2 per step. This preserves more information, enhancing U-Net's ability to reconstruct spatial details and improving segmentation accuracy by an average of 3.8% in Dice scores.
188. Prefer-DAS: Learning from Local Preferences and Sparse Prompts for Domain Adaptive Segmentation of Electron Microscopy
Core Problem: Unsupervised domain adaptation (UDA) strategies for delineating intracellular structures from large-scale electron microscopy (EM) often demonstrate limited and biased performance, hindering practical applications, and there's a need for more realistic and annotation-efficient settings using weak labels.
Key Innovation: Introduces Prefer-DAS, a promptable multitask model that pioneers sparse promptable learning and local preference alignment for domain adaptive segmentation. It integrates self-training and prompt-guided contrastive learning, allowing for flexible use of point prompts and introducing Local/Sparse/Unsupervised Preference Optimization (LPO/SLPO/UPO) for alignment with human feedback, achieving superior performance in both automatic and interactive segmentation modes.
189. A Statistical Approach for Modeling Irregular Multivariate Time Series with Missing Observations
Core Problem: Irregular multivariate time series with missing values pose significant challenges for predictive modeling, with deep learning approaches often being complex and computationally intensive.
Key Innovation: Proposes a simpler, effective statistical approach that extracts time-agnostic summary statistics (mean, std dev of observed values and changes) from irregular time series with missing data, achieving state-of-the-art performance with reduced computational complexity.
190. Fore-Mamba3D: Mamba-based Foreground-Enhanced Encoding for 3D Object Detection
Core Problem: Existing Mamba-based 3D object detection methods inefficiently encode abundant background information and suffer from response attenuation and restricted context representation when attempting foreground-only encoding.
Key Innovation: Proposes Fore-Mamba3D, a novel Mamba-based encoder that enhances foreground encoding by using a regional-to-global slide window for information propagation and a semantic-assisted and state spatial fusion module to enrich contextual representation, leading to superior 3D object detection.
191. VGGT-MPR: VGGT-Enhanced Multimodal Place Recognition in Autonomous Driving Environments
Core Problem: Existing multimodal place recognition (MPR) methods in autonomous driving rely on hand-crafted fusion strategies and heavily parameterized backbones, leading to costly retraining and limited robustness to environmental changes.
Key Innovation: VGGT-MPR, a multimodal place recognition framework that uses a Visual Geometry Grounded Transformer (VGGT) for both global retrieval and a training-free re-ranking mechanism, enhancing discriminative multimodal features and achieving robust performance in autonomous driving environments.
192. RAP: Fast Feedforward Rendering-Free Attribute-Guided Primitive Importance Score Prediction for Efficient 3D Gaussian Splatting Processing
Core Problem: Existing 3D Gaussian Splatting methods generate many redundant primitives, and current importance estimation relies on slow, rendering-based analyses that lack scalability and generalization.
Key Innovation: Proposes RAP, a fast, feedforward, rendering-free method that predicts primitive importance scores directly from intrinsic Gaussian attributes and local neighborhood statistics, improving efficiency and generalization for 3D scene reconstruction.
193. Multimodal Dataset Distillation Made Simple by Prototype-Guided Data Synthesis
Core Problem: Multimodal learning relies on large, costly datasets, and existing dataset distillation methods are often architecture-dependent, require full-dataset training, and struggle with very small subsets.
Key Innovation: Proposes a learning-free dataset distillation framework using CLIP to extract aligned image-text embeddings, obtain prototypes, and synthesize images with an unCLIP decoder, enabling efficient, scalable, and architecture-agnostic distillation.
194. Drift Localization using Conformal Predictions
Core Problem: Existing drift localization methods, relying on local testing schemes, tend to fail in high-dimensional, low-signal settings, making it difficult to determine which samples are affected by concept drift over time.
Key Innovation: A fundamentally different approach based on conformal predictions is proposed to localize concept drift, demonstrating improved performance on state-of-the-art image datasets by addressing the shortcomings of common local testing schemes.
195. Unlearning Noise in PINNs: A Selective Pruning Framework for PDE Inverse Problems
Core Problem: Physics-informed neural networks (PINNs) are highly sensitive to noise in observational data for PDE inverse problems, leading to distorted internal representations, impaired accuracy, and destabilized training.
Key Innovation: P-PINN, a selective pruning framework that 'unlearns' the influence of corrupted data in pretrained PINNs by evaluating a joint residual-data fidelity indicator and a bias-based neuron importance measure, identifying and removing noise-sensitive neurons to improve robustness and accuracy.
196. HeatPrompt: Zero-Shot Vision-Language Modeling of Urban Heat Demand from Satellite Images
Core Problem: Most municipalities lack detailed building-level data for accurate heat-demand maps, which are crucial for decarbonizing space heating.
Key Innovation: HeatPrompt, a zero-shot vision-language energy modeling framework that estimates annual heat demand using semantic features extracted from satellite images, basic GIS, and building-level features. It leverages pretrained Large Vision Language Models (VLMs) with a domain-specific prompt to extract visual attributes corresponding to thermal load.
197. RDBLearn: Simple In-Context Prediction Over Relational Databases
Core Problem: Many real-world prediction tasks reside in relational databases, where predictive signals are spread across multiple linked tables, making it challenging to apply existing tabular in-context learning (ICL) methods.
Key Innovation: RDBLearn, a simple recipe and toolkit that extends tabular ICL to relational prediction by automatically featurizing target rows using relational aggregations, materializing an augmented table, and applying an off-the-shelf tabular foundation model, outperforming strong supervised baselines.
198. Multiclass Calibration Assessment and Recalibration of Probability Predictions via the Linear Log Odds Calibration Function
Core Problem: Existing multicategory recalibration methods are limited in assessing single model calibration, often require under-the-hood model access, and provide output that is difficult for human analysts to understand.
Key Innovation: Multicategory Linear Log Odds (MCLLO) recalibration, a method that includes a likelihood ratio hypothesis test to assess calibration, does not require under-the-hood model access, and provides easily interpretable output, demonstrated effectively across various real-world classification problems.
199. Prognostics of Multisensor Systems with Unknown and Unlabeled Failure Modes via Bayesian Nonparametric Process Mixtures
Core Problem: Most existing prognostic models assume a fixed, known set of failure modes with labeled historical data, limiting their use for predictive maintenance in modern manufacturing systems where multiple, unpredictable, unknown, or unlabeled failure modes may emerge.
Key Innovation: A novel Bayesian nonparametric framework is proposed that unifies a Dirichlet process mixture for unsupervised failure mode discovery with a neural network-based prognostic module. An iterative feedback mechanism jointly learns and updates these modules to dynamically infer, expand, or merge failure modes as new data arrive, providing high prognostic accuracy and robust online adaptation.
200. Vid2Sid: Videos Can Help Close the Sim2Real Gap
Core Problem: Calibrating robot simulator physics parameters to match real hardware is difficult, often done manually or with black-box optimizers that lack interpretability, especially with limited external sensing.
Key Innovation: Vid2Sid, a video-driven system identification pipeline, couples foundation-model perception with a VLM-in-the-loop optimizer to analyze paired sim-real videos, diagnose concrete mismatches, and propose physics parameter updates with natural language rationales, achieving better accuracy and interpretability.
201. Cost-Aware Diffusion Active Search
Core Problem: Active search for recovering objects of interest in partially observable environments requires balancing exploration and exploitation, but lookahead algorithms typically rely on computationally expensive search trees.
Key Innovation: Leverages the sequence modeling abilities of diffusion models to sample cost-aware lookahead action sequences for active search, outperforming standard baselines in full recovery rate and computational efficiency compared to tree search methods.
202. Sculpting the Vector Space: Towards Efficient Multi-Vector Visual Document Retrieval via Prune-then-Merge Framework
Core Problem: State-of-the-art multi-vector Visual Document Retrieval (VDR) suffers from prohibitive overhead, and current efficiency methods (pruning, merging) create a difficult trade-off between compression rate and feature fidelity.
Key Innovation: Introduces "Prune-then-Merge," a novel two-stage framework that adaptively prunes low-information patches and then hierarchically merges the refined embeddings, consistently outperforming existing methods by extending near-lossless compression and providing robust performance at high compression ratios.
203. Goal-Oriented Influence-Maximizing Data Acquisition for Learning and Optimization
Core Problem: Active data acquisition for deep neural networks is challenging because most approaches rely on predictive uncertainty estimates that are difficult to obtain reliably.
Key Innovation: Proposes Goal-Oriented Influence-Maximizing Data Acquisition (GOIMDA), an active acquisition algorithm that avoids explicit posterior inference while remaining uncertainty-aware through inverse curvature, consistently reaching target performance with substantially fewer labeled samples or function evaluations than uncertainty-based baselines.
204. Denoising Particle Filters: Learning State Estimation with Single-Step Objectives
Core Problem: Learning-based state estimation in robotics often treats the problem as sequence modeling, leading to difficult-to-interpret models and expensive training due to unrolling sequences of predictions.
Key Innovation: Proposes a novel particle filtering algorithm where measurement models are learned implicitly by minimizing a denoising score matching objective from individual state transitions, offering competitive performance, composability, and reduced training cost compared to end-to-end methods.
205. Meta-Learning and Meta-Reinforcement Learning - Tracing the Path towards DeepMind's Adaptive Agent
Core Problem: Standard machine learning models struggle to adapt to novel tasks by utilizing prior knowledge, unlike humans, due to their reliance on task-specific training.
Key Innovation: Provides a rigorous, task-based formalization of meta-learning and meta-reinforcement learning, tracing the landmark algorithms that led to DeepMind's Adaptive Agent, consolidating essential concepts for generalist AI approaches.
206. Dirichlet Scale Mixture Priors for Bayesian Neural Networks
Core Problem: Specifying effective prior distributions in Bayesian Neural Networks (BNNs) is challenging, leading to issues like difficult interpretation, overconfident predictions, and vulnerability to adversarial attacks.
Key Innovation: Proposes a new class of prior distributions for BNNs, the Dirichlet scale mixture (DSM) prior, which induces structured, sparsity-inducing shrinkage, leading to sparser networks, improved adversarial robustness, and competitive predictive performance with fewer parameters.
207. Beyond Mimicry: Toward Lifelong Adaptability in Imitation Learning
Core Problem: Current imitation learning agents excel at replay but fail when contexts shift or goals evolve, indicating a foundational issue where the objective has been optimized for perfect mimicry rather than adaptability.
Key Innovation: Proposes a research agenda to redefine success in imitation learning from perfect replay to compositional adaptability, focusing on learning behavioral primitives and recombining them in novel contexts without retraining, establishing metrics and outlining interdisciplinary research directions.
208. Learning to See the Elephant in the Room: Self-Supervised Context Reasoning in Humans and AI
Core Problem: The mechanism by which humans and AI acquire contextual knowledge for scene understanding without explicit supervision remains unclear, limiting AI's ability to interpret scenes through relationships among elements.
Key Innovation: Introduces SeCo, a self-supervised model that learns contextual relationships from complex scenes using separate vision encoders and an external memory module, outperforming SOTA self-supervised methods and aligning with human contextual reasoning.
209. Revisiting Graph Neural Networks for Graph-level Tasks: Taxonomy, Empirical Study, and Future Directions
Core Problem: Evaluations of Graph Neural Networks (GNNs) for graph-level tasks are often limited by narrow datasets, task coverage, and inconsistent experimental setups, hindering generalizability and comprehensive understanding.
Key Innovation: Presenting a comprehensive experimental study and taxonomy of GNNs for graph-level tasks, proposing a unified evaluation framework (OpenGLT) to standardize assessment across diverse datasets and scenarios, and providing extensive empirical insights into GNN strengths and weaknesses.
210. VillageNet: Graph-based, Easily-interpretable, Unsupervised Clustering for Broad Biomedical Applications
Core Problem: Clustering large, high-dimensional datasets with diverse variables effectively, especially without prior knowledge of the optimal number of clusters.
Key Innovation: Developing "Village-Net," an unsupervised, graph-based clustering algorithm that autonomously determines the optimal number of clusters by first forming "villages" with K-Means and then applying a community detection algorithm (WLCF) on a weighted network of these villages, demonstrating competitive performance and computational efficiency.
211. Mamba-Based Graph Convolutional Networks: Tackling Over-smoothing with Selective State Space
Core Problem: Graph Neural Networks (GNNs) suffer from over-smoothing as model depth increases, causing node representations to become indistinguishable due to limitations in distinguishing the importance of information from different neighborhoods.
Key Innovation: Introducing MbaGCN, a novel Mamba-based graph convolutional architecture that tackles over-smoothing by adaptively aggregating neighborhood information through a Message Aggregation Layer, Selective State Space Transition Layer, and Node State Prediction Layer, providing greater flexibility and scalability for deep GNN models.
212. Are We Measuring Oversmoothing in Graph Neural Networks Correctly?
Core Problem: Existing metrics for oversmoothing in Graph Neural Networks (GNNs), such as Dirichlet energy, are unreliable and fail to reliably capture performance degradation in realistic scenarios, especially for typical GNN depths.
Key Innovation: Proposing to measure oversmoothing using the numerical or effective rank of feature representations, which consistently captures performance degradation in GNNs, even when traditional energy-based metrics fail.
213. Hier-COS: Making Deep Features Hierarchy-aware via Composition of Orthogonal Subspaces
Core Problem: Traditional classifiers treat labels as independent, failing to account for semantic hierarchies, while existing hierarchy-aware methods learn sub-optimal representations and are poorly assessed by current evaluation metrics.
Key Innovation: Introducing Hier-COS, a novel framework for unified hierarchy-aware fine-grained and multi-level classification that is theoretically consistent with the hierarchy, implicitly adapts learning capacity, and achieves state-of-the-art performance; also proposes HOPS, a new ranking-based evaluation metric.
214. Noise-Aware Generalization: Robustness to In-Domain Noise and Out-of-Domain Generalization
Core Problem: Existing methods for Learning with Noisy Labels (LNL) and Domain Generalization (DG) perform poorly when combined in Noise-Aware Generalization (NAG) settings, as DG methods are compromised by label noise and LNL methods overfit to easy domains.
Key Innovation: Proposes Domain Labels for Noise Detection (DL4ND), the first direct method for NAG, which leverages the observation that noisy samples show greater variation across domains, outperforming existing LNL and DG methods in combined settings.
215. Towards A Universal Graph Structural Encoder
Core Problem: Existing graph models struggle to capture and transfer structural information across different graph domains due to inherent topological differences and inadequate exploration of the graph embedding space.
Key Innovation: Proposes GFSE, a universal pre-trained graph encoder built on a Graph Transformer, using multiple self-supervised learning objectives to capture transferable, intricate multi-level, and fine-grained topological features across diverse graph domains.
216. Query-Based Adaptive Aggregation for Multi-Dataset Joint Training Toward Universal Visual Place Recognition
Core Problem: Deep learning methods for Visual Place Recognition (VPR) trained on single datasets suffer from dataset-specific inductive biases and limited generalization. Multi-dataset joint training, while promising, can lead to suboptimal performance due to divergences saturating feature aggregation layers.
Key Innovation: Proposes Query-based Adaptive Aggregation (QAA), a novel feature aggregation technique that uses learned queries as reference codebooks to enhance information capacity. This method generates robust descriptors, achieving balanced generalization across diverse datasets and maintaining peak performance comparable to dataset-specific models.
217. MoVieS: Motion-Aware 4D Dynamic View Synthesis in One Second
Core Problem: Efficiently reconstructing 4D dynamic scenes, synthesizing novel views, and tracking 3D points from monocular videos, while also enabling zero-shot applications like scene flow estimation and moving object segmentation.
Key Innovation: Presents MoVieS, a Motion-aware View Synthesis model that reconstructs 4D dynamic scenes from monocular videos in one second by representing scenes with pixel-aligned Gaussian primitives and explicitly supervising their time-varying motions, enabling unified modeling of appearance, geometry, and motion.
218. Collaborative Multi-Modal Coding for High-Quality 3D Generation
Core Problem: Existing 3D-native generative architectures often operate within single-modality paradigms or are restricted to 3D structures, overlooking the complementary benefits of multi-modality data (e.g., RGB, RGBD, point clouds) and limiting training dataset scope.
Key Innovation: TriMM, the first feed-forward 3D-native generative model that learns from basic multi-modalities through 'collaborative multi-modal coding' and auxiliary 2D/3D supervision, enabling the generation of superior quality 3D assets with enhanced texture and geometric detail using less training data.
219. VLM-Pruner: Buffering for Spatial Sparsity in an Efficient VLM Centrifugal Token Pruning Paradigm
Core Problem: Vision-language models (VLMs) have high computational costs due to numerous visual tokens, and existing pruning methods overlook inter-token redundancy and spatial relationships, leading to inefficient or spatially sparse token selections.
Key Innovation: VLM-Pruner, a training-free token pruning algorithm that explicitly balances redundancy and spatial sparsity using a centrifugal token pruning paradigm for near-to-far selection, a Buffering for Spatial Sparsity (BSS) criterion, a parallel greedy strategy, and selective fusion of salient information from discarded tokens.
220. A Conditional Generative Framework for Synthetic Data Augmentation in Segmenting Thin and Elongated Structures in Biological Images
Core Problem: The shortage of high-quality pixel-level annotated datasets for segmenting thin and elongated filamentous structures in biological images makes deep learning challenging.
Key Innovation: A conditional generative framework based on Pix2Pix, combined with a filament-aware structural loss, to generate realistic synthetic images of filaments, thereby augmenting data for improved segmentation performance.
221. PyraTok: Language-Aligned Pyramidal Tokenizer for Video Understanding and Generation
Core Problem: Existing discrete video VAE tokenizers learn visual codebooks at a single scale with limited vocabularies and shallow language supervision, leading to poor cross-modal alignment and zero-shot transfer in text-to-video generation and video understanding.
Key Innovation: PyraTok, a language-aligned pyramidal tokenizer that learns semantically structured discrete latents across multiple spatiotemporal resolutions using a novel Language aligned Pyramidal Quantization (LaPQ) module, achieving state-of-the-art video reconstruction, text-to-video quality, and zero-shot performance on various video understanding tasks.
222. GOT-Edit: Geometry-Aware Generic Object Tracking via Online Model Editing
Core Problem: Most generic object tracking (GOT) methods primarily rely on 2D features, neglecting 3D geometric cues, which makes them susceptible to challenges like partial occlusion, distractors, and variations in geometry and appearance.
Key Innovation: Proposes GOT-Edit, an online cross-modality model editing approach that integrates geometry-aware cues from a pre-trained Visual Geometry Grounded Transformer into a 2D generic object tracker. It uses null-space constrained updates to combine geometric information with semantic discrimination.
223. SAGE: Scalable Agentic 3D Scene Generation for Embodied AI
Core Problem: Existing 3D scene generation systems for embodied AI often rely on rule-based or task-specific pipelines, yielding artifacts and physically invalid scenes, and lack scalability for generating diverse, realistic, and simulator-ready environments.
Key Innovation: Presents SAGE, an agentic framework that automatically generates simulation-ready 3D environments given a user-specified embodied task. It couples multiple generators for layout and object composition with critics for semantic plausibility, visual realism, and physical stability, using iterative reasoning and adaptive tool selection.
224. Exploring Singularities in point clouds with the graph Laplacian: An explicit approach
Core Problem: There is a need for robust theoretical and methodological approaches to analyze the geometry of underlying manifolds in datasets, particularly for identifying and characterizing singularities within point clouds.
Key Innovation: Develops theory and methods using the graph Laplacian to analyze point cloud geometry, providing theoretical guarantees and explicit bounds on its functional forms near singularities. This enables the development of tests for singularities and methods to estimate their geometric properties.
225. A Mini-Batch Quasi-Newton Proximal Method for Constrained Total-Variation Nonlinear Image Reconstruction
Core Problem: High-quality image reconstruction using accurate nonlinear physical models is computationally demanding, and existing accelerated stochastic proximal methods (ASPM) are still expensive per iteration.
Key Innovation: Proposes a mini-batch quasi-Newton proximal method (BQNPM) tailored for image reconstruction with constrained total variation regularization. BQNPM requires fewer iterations to converge and includes an efficient approach to compute a weighted proximal mapping, demonstrating effectiveness and efficiency on 3D inverse-scattering problems.
226. The MAPS Algorithm: Fast model-agnostic and distribution-free prediction intervals for supervised learning
Core Problem: Computing reliable conditional prediction intervals in high-dimensional supervised learning is challenging, as existing methods often rely on restrictive assumptions, lack scalability, or only guarantee marginal coverage.
Key Innovation: The MAPS (Model-Agnostic Prediction Sets) algorithm, based on a novel lifted predictive model (LPM), which produces distribution-free conditional prediction intervals, adapts to any predictive model, scales to high-dimensional inputs, accounts for heteroscedastic errors, and offers asymptotic conditional coverage.
227. Optimizing High-Dimensional Oblique Splits
Core Problem: Optimizing high-dimensional oblique splits to enhance decision tree performance, especially for complex data-generating processes, faces computational challenges due to the potentially prohibitively large candidate split set.
Key Innovation: Establishing Sufficient Impurity Decrease (SID) convergence for s0-sparse oblique splits, demonstrating the trade-off between statistical accuracy and computational cost, and proposing progressive trees that optimize oblique splits through iterative refinement for integration into ensemble models, outperforming existing oblique tree models.
228. Feature Representation Transferring to Lightweight Models via Perception Coherence
Core Problem: Effectively transferring feature representation from large teacher models to lightweight student models, especially when the student model has weaker representational capacity and doesn't need to preserve the teacher's absolute geometry.
Key Innovation: A method for feature representation transfer based on a new notion of "perception coherence" and a corresponding loss function that considers dissimilarity ranking, allowing the student model to mimic the teacher's perception while preserving global coherence.
229. Bayesian Attention Mechanism: A Probabilistic Framework for Positional Encoding and Context Length Extrapolation
Core Problem: Existing positional encoding (PE) methods in Transformer-based language models lack theoretical clarity and rely on limited evaluation metrics for context length extrapolation claims.
Key Innovation: The Bayesian Attention Mechanism (BAM), a theoretical probabilistic framework that formulates positional encoding as a prior, unifies existing methods, and introduces a new Generalized Gaussian positional prior that significantly improves long-context generalization.
230. Esoteric Language Models: Bridging Autoregressive and Masked Diffusion LLMs
Core Problem: Masked Diffusion Models (MDMs) for language generation underperform autoregressive (AR) models in perplexity and lack key inference-time efficiency features like KV caching, despite offering parallel and controllable generation.
Key Innovation: Eso-LMs, a new family of models that fuses AR and MDM paradigms using causal attention, enabling exact likelihood computation and KV caching for MDMs for the first time, significantly improving inference efficiency and achieving state-of-the-art speed-quality for unconditional generation.
231. Probability Bounding: Post-Hoc Calibration via Box-Constrained Softmax
Core Problem: Modern neural networks often produce poorly calibrated probabilities despite high accuracy, leading to issues with underconfidence and overconfidence, which is a critical practical concern for decision-making.
Key Innovation: Probability bounding (PB), a novel post-hoc calibration method that mitigates underconfidence and overconfidence by learning lower and upper bounds on output probabilities, implemented via the box-constrained softmax (BCSoftmax) function with an efficient computation algorithm.
232. MIBoost: A Gradient Boosting Algorithm for Variable Selection After Multiple Imputation
Core Problem: Performing robust variable selection with statistical learning methods (like gradient boosting) when data has missing values and multiple imputation is used, as simple pooling strategies are suboptimal and sophisticated methods are hard to implement.
Key Innovation: Proposing MIBoost, a novel gradient boosting algorithm that extends the principle of defining a single loss function across imputed datasets to achieve a unified variable-selection mechanism, yielding comparable prediction performance to recent sophisticated methods.
233. Exact and Heuristic Algorithms for Constrained Biclustering
Core Problem: Effectively performing biclustering while incorporating prior knowledge through pairwise constraints, specifically for the k-densest disjoint biclique problem, which is computationally challenging.
Key Innovation: Developing both an exact tailored branch-and-cut algorithm based on a low-dimensional semidefinite programming (SDP) relaxation and an efficient heuristic based on low-rank factorization of the SDP, demonstrating superior performance on synthetic and real-world datasets.
234. Unfolded Laplacian Spectral Embedding: A Theoretically Grounded Approach to Dynamic Network Representation
Core Problem: Learning stable, consistent, and interpretable representations for dynamic relational data, particularly for normalized Laplacian operators where stability guarantees have been lacking.
Key Innovation: Introducing Unfolded Laplacian Spectral Embedding (ULSE), a principled extension that provides both cross-sectional and longitudinal stability guarantees under a dynamic stochastic block model, and a dynamic Cheeger-type inequality linking the spectrum to worst-case conductance.
235. Neurosymbolic Retrievers for Retrieval-augmented Generation
Core Problem: Traditional Retrieval Augmented Generation (RAG) systems lack transparency, interpretability, and debuggability due to opaque neural components, hindering trust in high-stakes domains.
Key Innovation: Introducing Neurosymbolic RAG, which integrates symbolic reasoning (knowledge graphs) with neural retrieval to enhance transparency and performance, proposing methods like Knowledge Modulation Aligned Retrieval (MAR), KG-Path RAG, and Process Knowledge-infused RAG, demonstrated in mental health risk assessment.
236. WAKESET: A Large-Scale, High-Reynolds Number Flow Dataset for Machine Learning of Turbulent Wake Dynamics
Core Problem: The critical scarcity of large, diverse, and high-fidelity datasets, particularly for high-Reynolds number turbulent flows, which hinders the application of machine learning to computational fluid dynamics for real-world engineering problems.
Key Innovation: Introduction of WAKESET, a novel, large-scale CFD dataset of highly turbulent flows (1,091 RANS simulations, augmented to 4,364 instances, up to Re 1.09 x 10^8) focused on underwater vehicle interactions, designed to enable the development and benchmarking of ML models for flow field prediction and surrogate modeling.
237. Interpretable Failure Analysis in Multi-Agent Reinforcement Learning Systems
Core Problem: The lack of developed methods for interpretable failure detection and attribution in Multi-Agent Reinforcement Learning (MARL) systems, especially critical for their deployment in safety-critical domains.
Key Innovation: A two-stage gradient-based framework for interpretable failure analysis in MARL, enabling accurate detection of the initial failure source (Patient-0), validation of domino effects, and tracing of failure propagation through learned coordination pathways, achieving high detection accuracy with interpretable geometric evidence.
238. A PolSAR Bridge Detection Method Integrating Cross-Sectional Probability Modeling and Graph Topology Analysis
Core Problem: Accurate bridge detection in polarimetric SAR (PolSAR) imagery is challenging due to bridge diversity, strong speckle noise, complex backgrounds, and missed detection for bridges spanning narrow river branches.
Key Innovation: A bridge detection method that combines cross-section probability modeling and graph topology analysis, which novelly constructs water networks by extracting and connecting cross-sections at water branch termini, enabling reliable detection of land regions connecting adjacent water network branches, achieving superior performance on Gaofen-3 and RADARSAT-2 datasets.
239. Prediction of thin sand body reservoirs using facies-model-constrained stochastic optimization inversion
Core Problem: Current seismic inversion techniques struggle to accurately characterize the distribution patterns of subtle, thin sand body reservoirs in the subsurface.
Key Innovation: Proposed a stochastic optimization inversion method for thin sand bodies based on facies-model constraints, incorporating dynamic forward modeling, multi-scale seismic data representation, and sedimentary facies analysis, demonstrating improved prediction accuracy for thin sand bodies.
240. Climate-induced internal displacement in the Southwestern Coast of Bangladesh: socioeconomic vulnerabilities and spatiotemporal dynamics
Core Problem: Climate change impacts (food insecurity, water scarcity, livelihood limitations, environmental degradation, salinity intrusion) are causing internal displacement and increasing socioeconomic vulnerabilities in coastal Bangladesh.
Key Innovation: A mixed-method approach combining household surveys (Livelihood Effect Index) and geospatial analysis (NDVI) to assess household-level vulnerability and spatiotemporal environmental dynamics, identifying key predictors of displacement and proposing policy priorities for resilience.
241. Modelling water use in Nepal’s highlands: a multidisciplinary and probabilistic framework
Core Problem: Mountain communities in Nepal face increasing vulnerability due to climate-induced shifts in water availability, necessitating integrated frameworks to assess water use patterns and identify vulnerable populations for targeted adaptation.
Key Innovation: Presents a multidisciplinary and probabilistic framework combining hydrological source analysis (isotopic/geochemical tracers in Bayesian mixing model) with socio-demographic survey data (Latent Class Analysis) to quantify seasonal water contributions, characterize community-level water use, and inform climate resilience strategies in high-altitude regions.
242. Road network robustness of financial centre cities: Rich but not effective
Core Problem: Understanding the robustness and 'robustness effectiveness' of road networks in financial center cities, particularly identifying critical intersections and the dichotomy between economic prominence and infrastructural robustness, is not well-understood.
Key Innovation: An investigation into 'robustness effectiveness' (comparison to random networks) of road networks in top-ranking cities, discovering that critical intersections are often peripheral, and revealing correlations between city robustness rankings and GFCI rankings.
243. Proactive planning for water distribution system rehabilitation and expansion under long-term uncertainty
Core Problem: Long-term planning of water distribution system (WDS) rehabilitation is challenged by multiple sources of uncertainty and the limitations of traditional approaches that assume pre-defined intervention timing, leading to inefficient decisions and reduced hydraulic reliability.
Key Innovation: A proactive planning framework consisting of identifying optimal rehabilitation timing through system performance evaluation and determining optimal rehabilitation actions through optimization, leading to more flexible, cost-effective, and reliable WDS rehabilitation plans under uncertainty.
244. Coastal to deep-marine geomorphic classification: A standardised framework and review of terms to support globally consistent mapping
Core Problem: The interdisciplinary nature of marine and coastal geomorphology leads to terminological nuances that hinder consistent mapping and evidence-based decision-making across disciplinary boundaries and scales.
Key Innovation: This study develops a comprehensive, two-part classification framework (morphological mapping and geomorphic interpretation) with standardized terms and a hierarchical structure, supported by open-access digital vocabularies and GIS tools, enabling consistent global-to-local scale geomorphic mapping.
245. WHU-STree: A multi-modal benchmark dataset for street tree inventory
Core Problem: Traditional street tree inventories are time-consuming and labor-intensive, and existing automated datasets are limited in scale, annotation, or modality, hindering comprehensive analysis for urban tree management.
Key Innovation: Introduced WHU-STree, a cross-city, richly annotated, multi-modal urban street tree dataset integrating synchronized point clouds and high-resolution images, supporting over 10 tasks for street tree inventory.
246. Smart spherical tracking probe for visualizing soil flow in cutter chamber of shield tunneling machine
Core Problem: Direct visualization and tracking of soil flow characteristics in the intricate slurry chamber environment of shield tunneling machines is challenging, hindering understanding of construction efficiency and safety.
Key Innovation: Development of a smart spherical tracking probe based on inertial navigation technology, capable of reconstructing representative soil flow trajectories with sufficient accuracy, offering a new approach for understanding internal flow dynamics in shield tunneling applications.
247. How does gross primary production uncertainty impact evapotranspiration prediction within the carbon–water coupled model?
Core Problem: Quantifying the impact of significant uncertainties in remote sensing-based Gross Primary Production (GPP) estimates on the accuracy and spatiotemporal patterns of Evapotranspiration (ET) predictions within carbon-water coupled models.
Key Innovation: A quantitative assessment using a classical carbon-water coupled modeling framework driven by 11 state-of-the-art RS-based GPP products and observations, demonstrating that GPP uncertainty significantly affects ET estimation accuracy (increasing RMSE by 26.94%, decreasing R2 by 13.58%) and leads to highly divergent spatiotemporal ET patterns.
248. Quantifying the drivers of river thermal regimes in the Hanjiang River Basin under climate change and reservoir construction
Core Problem: Quantifying and disentangling the respective impacts of climate change and reservoir construction on river water temperature (RWT) variations across different temporal and spatial scales in regulated river basins.
Key Innovation: Development of an integrated SWAT-HHO-LSTM modeling framework (coupling SWAT, an improved Harris Hawks Optimization-enhanced Long Short-Term Memory network, and CMIP6 climate projections) to perform attribution analysis, revealing a distinct temporal intensification of climate impacts and a spatial transition from dam-dominated RWT regulation upstream to climate-driven warming downstream in the Hanjiang River Basin.
249. Winter heat storage and thermal transport in the source reservoir of water diversion project
Core Problem: Inter-basin water diversion projects alter thermal boundary conditions of source reservoirs, leading to winter freezing risks in long-distance water conveyance systems, and there's a need for efficient utilization of reservoir thermal potential.
Key Innovation: Employed a 3D hydrodynamic-thermal coupled model to simulate winter thermal conditions in a source reservoir, revealing inter-reservoir heat budget differences and proposing a dual-objective "water–heat" regulation strategy (high water level for heat storage, high intake discharge for enhanced transfer, low outflow) to increase intake temperature and mitigate freezing risks.
250. Monitoring snow depth with ICESat-2 at mid-latitudes: a synergistic multi-scale framework integrating ground-airborne-satellite observations in northern Xinjiang, China
Core Problem: Accurately monitoring snow depth (SD) using ICESat-2 in mid-latitude regions, which is challenging due to the lack of large-scale, high-accuracy snow-free DEMs and ground-based validation.
Key Innovation: Established a 'ground-airborne-satellite' synergistic observation framework, using UAV-LiDAR as an intermediate-scale bridge, to validate ICESat-2 SD retrieval in northern Xinjiang, demonstrating its feasibility and robust overall performance while highlighting the critical influence of complex terrain on retrieval uncertainty.
251. Surrogate modeling for three-phase flow in porous media based on a temporal-attention-enhanced multiple-input operator network
Core Problem: Reducing the substantial computational costs associated with conventional numerical simulations of three-phase flow in porous media, which is crucial across various scientific and engineering fields.
Key Innovation: Proposed a temporal-attention-enhanced multiple-input operator network (TA-MIONet) for efficient surrogate modeling of three-phase flow, integrating temporal attention, a classifier/mask layer for discontinuities, and physics-informed constraints, demonstrating improved accuracy and efficiency for complex heterogeneous reservoirs compared to traditional methods.