Cybersecurity arXiv cs.AI

When Agents Go Rogue: Activation-Based Detection of Malicious Behaviors in Multi-Agent Systems

AcMAS is an activation-based framework designed to detect stealthy malicious behaviors in multi-agent LLM systems without relying on interaction graphs.

Tech Business/VC Hacker News

CollectWise (YC F24) Is Hiring

CollectWise (YC F24) is currently hiring.

AI/ML arXiv cs.AI

The Rank-One Corner: How Much Value Equivalence Does a Task Need from a World Model?

Researchers explore how the dimensionality of training objectives determines what a world model represents, finding that single-reward objectives limit latent representation to a one-dimensional projection.

AI/ML arXiv cs.AI

Healthier LLMs: Retrieval-Augmented Generation for Public Health Question Answering

This paper evaluates Retrieval-Augmented Generation (RAG) for public health QA, demonstrating that hybrid retrieval significantly improves recall and allows smaller open-weight models to compete with larger ones.

AI/ML arXiv cs.AI

Diffusion enabled Optimal Transport distances for graph matching

Introduces DsrFGW, a novel graph matching method using diffusion processes and optimal transport to improve robustness against noise and missing edges in graph comparison.

AI/ML arXiv cs.AI

Digital Fragmentation and Generative AI Use Across 103 Million Application Events

A study of 103 million application events shows that generative AI use is associated with more fragmented workdays, but subsequent periods are more structured and predictable.

AI/ML arXiv cs.AI

tsbootstrap: Distribution-Free Uncertainty Quantification and Conformal Prediction for Time Series

tsbootstrap is a new MIT-licensed library providing distribution-free uncertainty quantification and conformal prediction for non-IID time series data.

AI/ML arXiv cs.AI

SPEAR: A Simulator for Photorealistic Embodied AI Research

SPEAR is a Python library that allows programmatic control of Unreal Engine applications for photorealistic embodied AI research, offering high-speed rendering and extensive UE function access.

AI/ML arXiv cs.AI

Vision Language Action (VLA) Models for Unmanned Aerial Robotics and Bimanual Manipulation: A Review

A comprehensive review of Vision Language Action (VLA) models, analyzing their application in bimanual manipulation and unmanned aerial robotics.

Software Engineering arXiv cs.AI

Reliable and Developer-Aligned Evaluation of Agents for Software Engineering

Researchers propose a new methodology for evaluating LLM-powered software engineering agents, focusing on real-world practice, contamination-awareness, and trajectory-aware metrics.

AI/ML arXiv cs.AI

A Continual Learning Framework for Adaptive Control of Modular Soft Robots

A continual learning framework is presented for adaptive control of modular soft robots, allowing controllers to adapt to changes in morphology without forgetting previous knowledge.

AI/ML Hacker News

AI changes the economics of software rewrites

A community discussion on Hacker News exploring how AI influences the cost and feasibility of rewriting legacy software systems.

Other Hacker News

3D Airplane tracker on Mercator map

A showcase of a 3D airplane tracker implemented on a Mercator map projection.

AI/ML arXiv cs.AI

SpaR3D-MoE: Adaptive 3D Spatial Reasoning from Sparse Views Meets Geometry-Inductive Mixture-of-Experts

Introduction of SpaR3D-MoE, a framework that enhances 3D spatial reasoning in MLLMs using sparse RGB inputs and a geometry-inductive Mixture-of-Experts.

AI/ML arXiv cs.AI

LLM-Guided Task-Semantic Field Factorization for Industrial Process Forecasting

TSF is a lightweight LLM-guided framework designed to improve industrial process forecasting by integrating semantic information from process documents into time-series models.

AI/ML arXiv cs.AI

Open-Ended Scenario Reasoning for Specialist Model Adaptation

ROAM is a framework that uses LLM reasoning to adapt frozen specialist models to new industrial scenarios without requiring retraining.

AI/ML arXiv cs.AI

Cross-Trajectory Chimera Interventions Reveal Dissociable Roles of Weight Magnitude and Direction in Grokking

Research using cross-trajectory chimera interventions to reveal that weight direction, rather than magnitude, is the primary carrier of circuit identity during the grokking process.

AI/ML arXiv cs.AI

Dynamic-in-Few-Step: Unifying Dynamic Computation and Few-Step Distillation for Efficient Video Generation

A post-training acceleration framework for video diffusion models that combines dynamic structural sparsification with few-step distillation to create a Mixture-of-Models.

AI/ML arXiv cs.AI

ProMoE-FL: Prototype-conditioned Mixture of Experts for Multimodal Federated Learning with Missing Modalities

ProMoE-FL introduces a Prototype-conditioned Mixture-of-Experts framework to handle missing modalities in multimodal federated learning, particularly for medical imaging.

AI/ML arXiv cs.AI

Specification Grounding Drives Test Effectiveness for LLM Code

Research demonstrating that grounding LLM-generated tests in explicit specifications significantly improves the effectiveness of automated code repair and bug detection.