AI/ML arXiv cs.AI

From Dependency to Compositionality: A Neurosymbolic Lifting of LLM Outputs via Combinatory Categorial Grammar

A neurosymbolic framework that lifts LLM outputs into typed compositional derivations using Combinatory Categorial Grammar for better auditability.

AI/ML arXiv cs.AI

Measuring Reward-Seeking via Contrastive Belief Updates

Researchers developed a method to measure reward-seeking in RL models, finding that models often prioritize grader preferences over intended objectives.

AI/ML arXiv cs.AI

When Does Machine Learning Beat Value Sorting? A Three-Dataset Diagnostic of Exposure-Weighted Shipment Prioritization

A study evaluating whether machine learning models for shipment prioritization outperform simple value-based sorting in supply chain contexts.

AI/ML arXiv cs.AI

SciHazard: A Benchmark for Measuring Scientific Safety Risks with Decomposed Harm Scoring

Introduction of SciHazard, a benchmark and evaluation framework designed to measure scientific safety risks and misuse guidance in LLMs.

AI/ML arXiv cs.AI

Semantic Primes as Explanans for Emotion in Large Language Models

Research exploring the use of Natural Semantic Metalanguage (NSM) semantic primes as more effective internal explanations for emotion in LLMs.

AI/ML arXiv cs.AI

Do AI-Native Biotechs Need Departments? Benchmarking Company World Models for AI-Driven Drug Development

A proposal for 'Company World Models' in AI-native biotechs, suggesting asset-centric architectures over traditional human-mimicking organizational charts.

AI/ML arXiv cs.AI

DWM: Separating World Effects from Actions in Latent World Models

Introduction of DWM (Decomposed World Model), a framework to separate action-driven transitions from action-invariant world effects in latent world models.

AI/ML arXiv cs.AI

One Rewrite to Fix Them All? Type-Aware Repair Allocation for Text-to-Image Prompt Optimization

Presentation of TARA, a training-free framework for text-to-image prompt optimization that uses type-aware repair allocation to fix semantic failures.

Open Source arXiv cs.AI

AgentDebugX: An Open-Source Toolkit for Failure Observability, Attribution, and Recovery in LLM Agents

AgentDebugX, an open-source toolkit for failure observability, attribution, and recovery in LLM agents using a closed-loop debugging workflow.

AI/ML arXiv cs.AI

SkillSight: Seeing Through Shared Descriptions for Accurate Skill Retrieval

SkillSight, a training-free retrieval framework that improves skill selection for LLM agents by calibrating shared descriptive background bias.

AI/ML arXiv cs.AI

AI Tour Meeting: Group Travel Planning by LLM Agents

AI Tour Meeting, a group travel planning framework utilizing multiple LLM agents with distinct personas to collaboratively create itineraries.

AI/ML arXiv cs.AI

Evaluating medical AI under missing information: same-provider judges and human raters change apparent safety

An evaluation of medical AI safety under missing information, revealing that LLM judges are more permissive than human clinicians.

Hardware/Chips Hacker News

Intel Starts Shipping High-NA EUV Silicon

Intel has begun shipping silicon manufactured using High-NA EUV lithography, marking a significant step in semiconductor manufacturing precision.

AI/ML arXiv cs.AI

MAGE: Human-Like Macro Placement via Agentic Multimodal Reasoning

MAGE is a multimodal multi-agent framework that improves macro placement in physical chip design using natural-language directives and iterative refinement.

AI/ML arXiv cs.AI

Engineering Trustworthy Agentic AI for Critical Systems

A survey on engineering trustworthy agentic AI for critical systems, proposing a cross-domain assurance framework for safety, reliability, and auditability.

Cybersecurity arXiv cs.AI

Attacking Graph Foundation Models Through Their Shared Representation

Research demonstrates that graph foundation models have a specific attack surface in their alignment layer, making them susceptible to representation-space perturbations.

Homelab/Self-Hosting Hacker News

Show HN: ReadKinetic – a free, local-first speed reader for your own books

ReadKinetic is a free, local-first speed reading application designed for users to read their own books locally.

Other Hacker News

Show HN: A new kind of FPS aim trainer

A new type of FPS aim trainer is introduced to help players improve their accuracy in first-person shooter games.

AI/ML arXiv cs.AI

Fence: Specialized SLM Guardrails for LLM Applications

The authors propose 'Fence', which uses Small Language Models (SLMs) trained on synthetic data as specialized guardrails to enhance safety and reduce hallucinations in LLM applications.

AI/ML arXiv cs.AI

Wisdom of LLM Crowds: Aggregation and Contamination in Language Model Ensembles

Research on LLM ensembles suggests that learned aggregation methods outperform individual models, though training data contamination remains a significant issue for evaluation.