Global News Digest

Technology

arXiv

GIRL-DETR: Gradient-Isolated Reinforcement Learning for Video Moment Retrieval

GIRL-DETR uses gradient-isolated RL to optimize non-differentiable metrics in lightweight video moment retrieval, preserving features while boosting precision on Charades-STA, QVHighlights, and TACoS.

arXiv

Behavior-Invariant Task Representation Learning with Transformer-based World Models for Offline Meta-Reinforcement Learning

This paper introduces a framework combining behavior-invariant task representations with Transformer-based world models for offline meta-RL. It mitigates distribution shifts and prevents model exploitation, achieving superior generalization in sparse-reward settings.

arXiv

SkyShield: Occupancy as a Safety Interface for Low-Altitude UAV Autonomy

SkyShield introduces the first monocular semantic occupancy benchmark for low-altitude UAVs, featuring 36,000 CARLA samples and the KAR-mIoU metric. It also presents SkyOcc, a baseline integrating UAV dynamics for enhanced safety-aware navigation.

arXiv

Bayesian Inference of Nonlinear Malaria Dynamics in Ghana via an Ensemble Markov Chain Monte Carlo Sampler

This study uses Bayesian MCMC to model Ghana’s nonlinear malaria dynamics, achieving high accuracy. Projections indicate a 2024–2026 resurgence, supporting evidence-based control decisions.

arXiv

DASH: Dual-Branch Score Distillation for Guidance-Calibrated Compact Diffusion Models

DASH improves compact diffusion models by supervising both score branches, preventing degenerate solutions and preserving guidance fidelity. It achieves high-quality results with significant compression, outperforming training from scratch.

arXiv

Extending Causal Metamodeling to a non-Markovian Queue

This study extends causal metamodeling via modular dynamic Bayesian networks to non-Markovian queues using phase-type distributions. Experiments on G/M/1 queues show the approach achieves high accuracy and inference speeds orders of magnitude faster than direct simulation.

arXiv

Dynamic Coordination Strategy Selection for Enterprise Multi-Agent Systems

This study finds that while specific winning coordination strategies vary by model, predicted strategies consistently remain near-optimal across enterprise tasks. This validates a dynamic selection approach, though single-agent workflows may suffice for structured compliance.

arXiv

Certificates without Electrons? Theory and Evidence on Impacts from AI-Driven Power Demand

AI-driven power demand undermines grid reliability via a "timing wedge" in renewable credits, increasing emissions and prices. On-site generation mitigates these adverse effects by aligning consumption with supply.

arXiv

SkillPager: Query-Adaptive Intra-Skill Navigation via Semantic Node Retrieval

SkillPager optimizes LLM prompt efficiency by retrieving semantic nodes from skill documents, reducing token usage by 47% while maintaining high context sufficiency. It outperforms graph-based baselines by leveraging typed granularity over fixed chunks.

arXiv

Hybrid Probabilistic Forecasting of Under-Five Malaria Admissions in Ghana: A Gaussian Process Regression with Holt-Winters Smoothing

This study combines Gaussian Process Regression with Holt-Winters smoothing to forecast under-five malaria admissions in Ghana, significantly outperforming standalone methods. The hybrid model provides robust, probabilistic district-level projections for early warning systems.

arXiv

Beyond Independent Manipulation: Individual Fairness-aware Strategic Classification with Peer Imitation

IFSC models strategic classification where agents imitate peers for individual fairness. It uses robust learning to improve fairness consistency and reduce manipulation distortions.

arXiv

From Cues to Horizons: Dynamic Risk Horizon Profiling for Trajectory Prediction

The authors propose Risk Horizon Profiling, a dynamic risk-aware module for trajectory prediction. It outperforms baselines on highway and urban datasets, significantly reducing prediction errors.

arXiv

RefDiffNet: Learning to Expose Subtle PCB Defects Before Detection

RefDiffNet enhances PCB defect detection by aligning inspected images with defect-free references to highlight subtle anomalies. It boosts mAP by up to 18% across various detectors with negligible computational overhead.

arXiv

MoEIoU: Rethinking Bounding-Box Regression as a Mixture of Experts

MoEIoU introduces a dynamic, mixture-of-experts loss for bounding-box regression, adapting focus from position to overlap during training. It outperforms state-of-the-art methods on COCO and VOC, improving accuracy and convergence.

arXiv

Task diversity produces systematic transfer but inhibits continual reinforcement learning

Task diversity enables systematic zero-shot transfer but inhibits continual reinforcement learning. High shift frequencies cause performance plateaus and catastrophic forgetting, limiting sustained adaptation.

arXiv

Benchmarks for Vision-Language Models in Urban Perception Should Be Reliability-Aware and Negotiated

This study argues that urban perception benchmarks for VLMs must prioritize reliability and treat labels as negotiable. It demonstrates that model alignment correlates with human judgment reliability, highlighting the need for transparent uncertainty handling.

arXiv

GenPT: Beyond Self-Report for Reliable LLM Psychometrics via Generative Projective Testing

GenPT replaces biased self-reports with generative projective testing for robust LLM psychometrics. It offers superior stability and sensitivity to psychological shifts compared to traditional questionnaires.

arXiv

Memory-Efficient LLM Training with Dynamic Sparsity: From Stability to Practical Scaling

SMET stabilizes memory-efficient LLM training via dynamic sparsity using optimizer warm-up and selective gradient storage. This enables stable, scalable sparse pre-training as a practical alternative to dense methods.

arXiv

Accuracy, Stability, and Repeated-Run Reliability of Large Language Models on Deterministic Programming Tasks

This study reveals that standard LLM accuracy metrics overstate performance by ignoring stability. Repeated-run analysis is essential for reliable evaluation of deterministic programming tasks.

arXiv

Dive into Waves: Morlet Spectral Transformer for Cross-Subject Emotion Decoding from EEG

The Morlet Spectral Transformer (MST) enhances cross-subject EEG emotion decoding via wavelet tokenization, drift removal, and frequency-specific spatial projection. It outperforms existing methods on SEED datasets without pretraining.