Global News Digest

Technology

arXiv

Pause and Think: A Dataset and Benchmark for Video-Grounded Assistive Action Suggestion

Pause-and-think-T dataset and benchmark enable a 4B-parameter model to match GPT-5.2’s scene understanding with 59x fewer parameters than Qwen3-VL, proving focused reasoning supervision allows smaller models to generalize effectively.

arXiv

Linguistics-Aware Non-Distortionary LLM Watermarking

LUNA is a linguistically adaptive watermarking framework achieving high detection accuracy (AUROC 0.9959) with minimal text distortion. It outperforms baselines across six languages by maintaining low perplexity shifts and preserving content quality.

arXiv

LP5X-PIM Sim: A High-Fidelity HW/SW Integrated Simulator for LPDDR5X-PIM

Samsung’s LP5X-PIM Sim is a high-fidelity HW/SW integrated simulator for LPDDR5X-PIM. It enables precise assessment of energy efficiency and performance optimization.

arXiv

LinguIUTics at PsyDefDetect: Iterative Imbalance-Aware Fine-tuning of Qwen3-8B for Psychological Defense Mechanism Classification

LinguIUTics ranked 4th in PsyDefDetect 2026 by fine-tuning Qwen3-8B with QLoRA and imbalance-aware strategies. This approach significantly boosted minority-class recall, improving macro F1 by 24.4% over the baseline.

arXiv

Demystifying the Optimal Fair Classifier in Multi-Class Classification

This study introduces a probabilistic framework and two algorithms to achieve the optimal accuracy-fairness trade-off in multi-class classification. Empirical results validate their effectiveness in balancing predictive performance with equitable treatment across demographic groups.

arXiv

MESA: Improving MoE Safety Alignment via Decentralized Expertise

MESA decentralizes safety in MoE LLMs via Optimal Transport, reallocating expertise to mitigate "Safety Sparsity." This approach robustly defends against attacks while preserving model helpfulness.

arXiv

Collaborative Few-Step Distillation and Low-Bit Quantization for Wan2.2 Dual-Expert Video Diffusion Models

This study optimizes Wan2.2 video models via joint few-step distillation and low-bit quantization. The approach balances quality and efficiency, outperforming baselines at 8 and 20 steps.

arXiv

Beyond the Mouth: Upper-Face Affective Cues in Audiovisual Sentence Recognition under Acoustic Uncertainty

Upper-face emotional cues enhance audiovisual sentence recognition robustness and model calibration under acoustic noise, despite minimal direct accuracy gains.

arXiv

Scaling Behavior of Single LLM-Driven Multi-Agent Systems

This study reveals that LLM-based multi-agent systems face diminishing returns due to coordination overhead, not context limits. Optimal performance requires robust base models and task-specific agent counts, challenging the assumption that more agents automatically yield better results.

arXiv

COPF: An Online Framework for Deployment-Stable Counterfactual Fairness in Evolving Graphs

COPF is an online framework ensuring stable counterfactual fairness in dynamic link recommendation systems. It mitigates exposure disparities via graph-aware estimators and multicalibration with minimal utility loss.

arXiv

The Paradox of Outcome Optimization: A Causal Information-Theoretic Bound on Reasoning Shortcuts in LLMs

The paper proves outcome-optimized LLMs collapse reasoning via spurious correlations. It shows process supervision acts as a topological filter, forcing models to learn robust causal mechanisms instead of shortcuts.

arXiv

Shape Your Body: Value Gradients for Multi-Embodiment Robot Design

This study uses frozen, multi-embodiment value functions as differentiable surrogates to optimize robotic designs via value gradients. This approach efficiently refines morphology and identifies performance constraints across diverse robot configurations.

arXiv

Information-Theoretic Lower Bounds for Bit-Constrained Stochastic Optimization via a Reduction to Compressed Gaussian Mean Estimation

This paper derives information-theoretic lower bounds for bit-constrained stochastic optimization by reducing it to compressed Gaussian mean estimation. It establishes unconditional communication and statistical limits, validated by a near-matching achievability result.

arXiv

Multi-Agent Conformal Prediction with Personalized Statistical Validity

PFWCP ensures personalized statistical validity in multi-agent conformal prediction by combining local density weighting with privacy-preserving quantile aggregation. It outperforms existing federated baselines in calibration quality while maintaining strict privacy and asymptotic validity guarantees.

arXiv

WaveFilter: Enhancing the Long-Context Capability of Diffusion LLMs via Wavelet-Guided KV Cache Filtering

WaveFilter enhances Diffusion LLMs’ long-context performance via training-free wavelet-guided KV cache filtering. It identifies pivotal tokens to build sparse caches, improving efficacy without retraining.

arXiv

EPIC: Efficient and Parallel Inference under CFG Constraints for Diffusion Language Models

EPIC accelerates CFG-constrained diffusion language model inference by 67.5% via parallel token commitment and optimized parsing, restoring the efficiency lost in prior methods.

arXiv

Quantum Tunneling-Aware Machine Learning: Physics-Derived Noise Models for Robust Deployment

QTAML uses physics-derived noise models to correct quantum tunneling errors in AI. Its TAC algorithm boosts accuracy with significantly less ECC overhead than baselines.

arXiv

SORA: Free Second-Order Attacks in Fast Adversarial Training

SORA prevents catastrophic overfitting in fast adversarial training via PertAlign and adaptive step-sizes. It achieves state-of-the-art robustness and clean accuracy with minimal computational overhead.

arXiv

Causal Density Functions

Causal density functions use Radon-Nikodym derivatives to contrast interventional and observational distributions, enabling pointwise causal effect estimation. This framework allows for empirical testing and scoring of directed influence via reweighted observational expectations.

arXiv

Logit Distillation on Manifolds: Mapping by Learning

This paper proposes a manifold-based logit distillation method using layer-wise projection and LoRA, reducing student parameters to under 1% while significantly improving word error rates compared to existing techniques.