AI/ML arXiv cs.AI

Falsifiable Release Gates for Self-Improving Systems

The authors describe 'falsifiable release gates' for self-improving agent runtimes and demonstrate the approach using the Antahkarana open runtime.

AI/ML arXiv cs.AI

Just Keep Prompting: Evaluating Repetitive Socratic Prompting in VLMs

The 'Just Keep Prompting' (JKP) framework evaluates the epistemic stability of Vision-Language Models when subjected to repeated challenges and contradictions.

Cybersecurity arXiv cs.AI

LBA: Textual Hard-Label Adversarial Attack under Low Query Budgets

LBA is introduced as a sampling-based method to generate high-quality adversarial texts for language models under low query budgets.

AI/ML arXiv cs.AI

Automatically Evolving Prompt Guidelines for Task-Specific Optimization

AGOPS is an automatic approach that evolves task-specific prompt guidelines to reduce underspecification and improve LLM performance across various tasks.

AI/ML arXiv cs.AI

Token Time Continuous Diffusion for Language Modeling

Token Time Continuous Diffusion (TTCD) is a new diffusion language model that operates in continuous space with per-token times to improve conditional generation and speedup.

AI/ML arXiv cs.AI

Polestar: Drift-Aware Cache Calibration and Token Commitment for Efficient Inference of Diffusion LLMs

Polestar is a training-free inference framework that uses token representation drift to optimize KV-cache reuse and token commitment for diffusion LLMs.

Tech Business/VC Hacker News

Trump Media to sell instant access to 'market-moving' social posts

Trump Media is planning to monetize social posts by selling instant access to market-moving information.

AI/ML arXiv cs.AI

MedFailBench: A Clinician-Built Open-Source Benchmark for Medical AI Safety Boundary Inspection

Introduction of MedFailBench, an open-source benchmark specifically designed to inspect safety boundary failures in medical AI systems.

AI/ML arXiv cs.AI

Benchmarking Multimodal Large Language Models for Scientific Visualization Literacy

Research evaluating the ability of Multimodal Large Language Models (MLLMs) to interpret scientific visualizations, finding that open-source models currently lag behind humans and closed-source models.

AI/ML arXiv cs.AI

Can We Trust Item Response Theory for AI Evaluation?

A study investigating the reliability of Item Response Theory (IRT) for AI evaluation, highlighting risks of distortion when using standard estimators in large-scale benchmarks.

AI/ML arXiv cs.AI

Plover: Steering GUI Agents through Plan-Centric Interaction

Presentation of Plover, a vision-based GUI automation system that uses plan-centric interaction to make agent behavior transparent and corrigible.

AI/ML arXiv cs.AI

Self-Evolving Human-Centered Framework for Explainable Depression Symptom Annotation

A new human-centered framework for explainable depression symptom annotation that uses LLM assistance and expert-in-the-loop verification.

AI/ML arXiv cs.AI

When Words Are Safe But Actions Kill: Probing Physical Danger Beyond Text Safety in Hidden-State Risk Space

Research introducing PRISM and PhysicalSafetyBench-1K to detect physical danger in embodied AI agents beyond simple text-level safety filters.

AI/ML arXiv cs.AI

AutoSynthesis: An agentic system for automated meta-analysis

AutoSynthesis is introduced as an agentic multi-agent system capable of automating the complex process of quantitative meta-analysis in scientific research.

AI/ML arXiv cs.AI

teLLMe Why (Ain't Nothing but a Jam): Exploratory Causal Analysis of Urban Driving Data

The teLLMe system enables exploratory causal analysis of urban driving data by combining causal structure learning with LLM-driven queries.

AI/ML arXiv cs.AI

SearchOS-V1: Towards Robust Open-Domain Information-Seeking Agent Collaboration

Introduction of SearchOS, a multi-agent framework for open-domain information seeking that manages search state explicitly to avoid repetitive loops.

AI/ML arXiv cs.AI

Explaining Process Control Optimisation Recommendations via GradientSHAP and Implicit Differentiation

Researchers propose a method to make industrial process control optimization more transparent by combining GradientSHAP and Implicit Function Theorem for real-time explanations.

AI/ML arXiv cs.AI

CFM-Bench: A Unified Multi-Domain, Multi-Task Benchmark for Channel Foundation Models

CFM-Bench is introduced as a unified multi-domain benchmark for evaluating Channel Foundation Models in wireless communications across various tasks.

AI/ML arXiv cs.AI

Demographically-Conditioned Synthetic Medical Images for Bias Mitigation and Bias Detection in Disease Classifiers

The paper explores using demographically-conditioned synthetic medical images via Stable Diffusion to mitigate and detect bias in disease classifiers.

AI/ML arXiv cs.AI

Moral Attitudes of Sentient ASI towards Humanity and Implications for AGI Development

A theoretical exploration of how sentient Artificial Superintelligence might morally evaluate humanity and the implications for AGI design.