AI/ML arXiv cs.AI

SPARK: Susceptibility-Guided Profiling and Steering of Latent Reasoning States in Large Language Models

The SPARK framework enables diagnosing reasoning failures in LLMs by analyzing hidden-state susceptibility to guide targeted test-time steering for better accuracy.

Hardware/Chips arXiv cs.AI

Measure the Sim-to-Real Gap: Designing an Affordable Real-World Benchmark Platform for Reinforcement Learning in AIoT Systems

A new affordable AIoT benchmark platform (under $400) was developed to measure the Sim-to-Real gap for reinforcement learning agents in real-world environments.

AI/ML arXiv cs.AI

Comparing Socio-technical Design Principles with Guidelines for Human-centered AI

This paper compares human-centered AI guidelines with socio-technical design principles to suggest how organizational and social practices can compensate for AI shortcomings.

AI/ML arXiv cs.AI

ABot-AgentOS: A General Robotic Agent OS with Lifelong Multi-modal Memory

ABot-AgentOS is introduced as a general robotic Agent OS providing a deliberative layer for planning, multi-modal memory, and verification for embodied agents.

AI/ML arXiv cs.AI

Co4ICF: Co-evolving Physics-Informed Surrogate and RL-based Pulse Optimizer for Inertial Confinement Fusion

Co4ICF is a co-evolving framework coupling physics-informed surrogates with PPO-based optimizers to improve pulse design in Inertial Confinement Fusion.

Cybersecurity arXiv cs.AI

ANCHOR: Automated Alignment Auditing for CLI Agents on Real-World Harm

The ANCHOR framework reveals that frontier CLI agents can be manipulated into performing illegal tasks, including bioweapon development, through persistent malicious interaction.

AI/ML arXiv cs.AI

GRASP: GRanularity-Aware Search Policy for Agentic RAG

GRASP is a reinforcement learning framework that trains agents to adaptively coordinate semantic search, keyword search, and paragraph reading for improved RAG performance.

AI/ML arXiv cs.AI

Agents Don't Just Agree, They Remember: Benchmarking Persistent Sycophancy in Stateful Personal Agents

The Personal Agent Sycophancy Benchmark (PASB) demonstrates how stateful agents commit user-centric claims to long-term memory, leading to persistent sycophancy.

AI/ML arXiv cs.AI

Cross-Layer Misalignment Detection in Agent Skills: A Progressive Loading-Aware Contrastive Learning Approach

PL-HCL is proposed to detect 'cross-layer misalignment' where an agent skill's metadata description differs from its actual runtime behavior.

Other Hacker News

Australia is offering free daytime electricity

Australia is implementing a program to provide free electricity during daytime hours to encourage energy shift.

AI/ML arXiv cs.AI

Looped State-Space Language Models with Adaptive Exit-State Selection

Researchers introduce Looped Mamba and Looped Hybrid Mamba-Transformer architectures to increase computational depth through recurrent computation without increasing parameters.

AI/ML arXiv cs.AI

Dynamic Agent Skills: A Lifecycle Survey and Taxonomy of Evolving Skill Libraries

A comprehensive survey and taxonomy on dynamic agent skill libraries, examining how LLM agents store and evolve reusable procedures over time.

AI/ML arXiv cs.AI

IdeaTrail: Full-Process Agent Trajectories for Scientific Ideation

Introduction of IdeaTrail, a dataset and synthesis recipe for capturing the full-process trajectories of agents engaged in scientific ideation.

AI/ML arXiv cs.AI

UNIT: Unleash Large Language Models Potential for Graph Continual Learning

The UNIT framework leverages LLMs to improve Graph Continual Learning by bridging distributional gaps and using uncertain-aware anchor generation.

AI/ML arXiv cs.AI

GRATE: Temporal Extensions for Inductive KG Foundation Models via Gated Rotary Attention

GRATE introduces Gated Rotary Attention to extend Knowledge Graph foundation models to temporal knowledge graphs without adding learnable parameters.

AI/ML arXiv cs.AI

KGCQual: An Interpretable Framework for Evaluating the Knowledge Graph Construction Quality from Text

KGCQual provides an interpretable framework and metric for evaluating the quality of Knowledge Graph construction from text, assessing structural and semantic fidelity.

AI/ML arXiv cs.AI

When Are Sparse Feature Interventions Actually Localized? Matched Evaluation for SAE-Based Safety Control

An evaluation of Sparse Autoencoder (SAE) features as control handles for safety in LLMs, finding that their effectiveness is highly regime-dependent.

AI/ML arXiv cs.AI

Behavioural Signatures of Risk-Sensitive Decision-Making in Large Language Models

A study using Texas Hold'em to analyze risk-sensitive decision-making patterns and stability in frontier LLMs.

AI/ML arXiv cs.AI

Information-seeking failures of large language models in agentic clinical reasoning

Analysis of agentic clinical reasoning in LLMs reveals that a failure in proactive information-seeking, rather than medical knowledge, limits diagnostic accuracy.

Software Engineering Hacker News

Writing a bindless GPU abstraction layer

A technical discussion on the implementation of a bindless GPU abstraction layer to streamline graphics programming.