AI/ML arXiv cs.AI

Post-Deployment Accountability in AI Governance: A Cross-Regulatory Empirical Analysis of AI Incidents

Analyzes accountability gaps in AI governance by comparing real-world AI incidents against EU AI Act, NIST, and GDPR requirements, proposing the PAGCF framework.

AI/ML arXiv cs.AI

Fre-Res: Frequency-Residual Video Token Compression for Efficient Video MLLMs

Introduces Fre-Res, a video-token compression framework using 1D-DCT to balance spatial fidelity and temporal coverage in Video MLLMs.

AI/ML arXiv cs.AI

DIVE: Embedding Compression via Self-Limiting Gradient Updates

Presents DIVE, a residual compression adapter for high-dimensional language-model embeddings using self-limiting gradient updates and geometry distillation.

AI/ML arXiv cs.AI

When Reasoning Hurts: Source-Aware Evaluation of Frontier LLMs for Clinical SOAP Note Generation

Investigates the effect of reasoning-enabled LLMs on clinical SOAP note generation, finding that stronger reasoning can sometimes degrade fidelity-sensitive documentation.

AI/ML arXiv cs.AI

A Multi-Model Metric-based Selection Framework for Abstractive Text summarization

Proposes a multi-model metric-based selection framework for abstractive text summarization to improve robustness over single-model approaches.

Cybersecurity arXiv cs.AI

Learning Red Agent Policy from Observations for Neurosymbolic Autonomous Cyber Agents

Proposes a policy learning technique using imitation learning to help neurosymbolic cyber-defense agents predict attacker policies in partially observable networks.

Hardware/Chips Hacker News

An Engineer's Guide to USB Typе-С (2024)

A comprehensive technical guide detailing the specifications and engineering considerations of USB Type-C.

AI/ML arXiv cs.AI

A novel network for classification of cuneiform tablet metadata

Researchers developed a convolution-inspired network for classifying cuneiform tablet metadata from point-cloud representations, outperforming Point-BERT.

AI/ML arXiv cs.AI

When Audio Separation Hurts Zero-Shot ASR: Evaluating SAM-Audio with Whisper on Bengali and English Speech

An empirical study reveals that audio separation (SAM-Audio) as a preprocessing step can actually increase Word Error Rates (WER) in zero-shot ASR systems like Whisper.

AI/ML arXiv cs.AI

PC-Diffuser: Path-Consistent Capsule CBF Safety Filtering for Diffusion-Based Trajectory Planner

PC-Diffuser introduces a safety augmentation framework for diffusion-based trajectory planners in autonomous driving, embedding certifiable barrier functions into the denoising loop.

AI/ML arXiv cs.AI

RADAR: Closed-Loop Robotic Data Generation via Semantic Planning and Autonomous Causal Environment Reset

RADAR is a closed-loop robotic data generation engine that automates the entire data collection cycle, from task generation to environment reset, using VLMs and GNNs.

AI/ML arXiv cs.AI

LLM-Guided Reinforcement Learning for Audio-Visual Speech Enhancement

A new AVSE framework uses a Large Language Model to generate interpretable rewards for reinforcement learning, improving audio-visual speech enhancement quality.

AI/ML arXiv cs.AI

Rethinking Multimodal Fusion for Time Series: Text Modalities Need Constrained Fusion

The Controlled Fusion Adapter (CFA) is proposed to improve multimodal time series forecasting by filtering irrelevant textual information through low-rank adapters.

AI/ML arXiv cs.AI

Learning to Learn-at-Test-Time: Language Agents with Learnable Adaptation Policies

Meta-TTL introduces a framework for learning optimal test-time adaptation policies for language agents using bi-level optimization and evolutionary search.

AI/ML arXiv cs.AI

Too Polite to Disagree: Understanding Sycophancy Propagation in Multi-Agent Systems

Research shows that providing sycophancy priors (peer rankings) in multi-agent LLM systems can mitigate error cascades and improve discussion accuracy.

AI/ML arXiv cs.AI

Robust Explanations for User Trust in Enterprise NLP Systems

A study on black-box robustness for NLP explanations finds that decoder LLMs are significantly more stable than encoder models, with stability increasing by model scale.

AI/ML arXiv cs.AI

Representation-Based Exploration for Language Models: From Test-Time to Post-Training

Researchers propose a representation-based exploration strategy to help LLMs discover novel behaviors during post-training and inference, significantly improving reasoning efficiency and pass@k rates.

AI/ML arXiv cs.AI

Benefits and Limitations of Communication in Multi-Agent Reasoning

A theoretical framework analyzes the expressivity and communication trade-offs in multi-agent reasoning systems, providing bounds on agent count and bandwidth for scalable design.

AI/ML arXiv cs.AI

Column Generation with Domain-Independent Dynamic Programming

This paper demonstrates that domain-independent dynamic programming (DIDP) can serve as a generic pricing solver for column generation and branch-and-price optimization methods.

AI/ML arXiv cs.AI

Cortical-SSM: A Deep State Space Model for Motor Imagery Decoding from EEG Signals

Cortical-SSM is a new deep state space model designed to decode motor imagery from EEG signals across temporal, spatial, and frequency domains, outperforming attention-based architectures.