Cybersecurity arXiv cs.AI

Agentra: A Supervisable Multi-Agent Framework for Enterprise Intrusion Response

Agentra is a multi-agent framework for enterprise intrusion response that converts alerts into structured plans grounded in MITRE and NIST standards.

Software Engineering Hacker News

Sogen – High-performance Windows and Linux userspace emulator

Sogen is a high-performance userspace emulator designed for both Windows and Linux environments.

Other Ars Technica

Hulk, Punisher join Peter Parker in Spider-Man: Brand New Day trailer

A new trailer for the Spider-Man: Brand New Day game has been released, featuring Hulk and Punisher.

AI/ML arXiv cs.AI

Vibe Coding Ate My Homework: An evaluation of AI approaches to greenfield software engineering and programming

Researchers evaluate 'vibe coding,' a paradigm where natural language prompts replace traditional code syntax for greenfield software engineering.

AI/ML arXiv cs.AI

A Link between Shock-wave Theory and Symmetry-reduced Stochastic Gradient Descent for Artificial Neural Networks

This paper establishes a mathematical link between shock-wave theory and symmetry-reduced stochastic gradient descent in artificial neural networks.

AI/ML arXiv cs.AI

Attribution-Guided and Coverage-Maximized Pruning for Structural MoE Compression

A new structural pruning framework for Mixture-of-Experts (MoE) models improves compression and memory footprint without significant accuracy loss.

AI/ML arXiv cs.AI

DRIFT: Refining Instruction Data via On-Policy Data Attribution

DRIFT is a data refinement method for SFT that uses on-policy rollouts to improve the performance ceiling of LLMs.

AI/ML arXiv cs.AI

TRIDENT: Breaking the Hybrid-Safety-Physics Coupling for Provably Safe Multi-Agent Reinforcement Learning

TRIDENT is a MARL framework designed for provably safe coordination in networked cyber-physical systems with hybrid actions and physics-governed dynamics.

AI/ML arXiv cs.AI

SAGE: Retain-Aware Post-Hoc Sanitization of Final Unlearning Vector

SAGE is a post-hoc sanitization method that reduces the trade-off between unlearning undesirable knowledge and preserving useful capabilities in LLMs.

Cybersecurity arXiv cs.AI

Conflict-Aware Retriever Editing for Knowledge Injection Attacks on LLM-Based RAG Systems

CAREATTACK introduces a model-centric retriever attack framework for injecting malicious knowledge into RAG systems via retriever editing.

AI/ML arXiv cs.AI

Ghost Attractor Networks: Basin-Structured Dynamical Decoders for Closed-Loop Sequential Generation

Ghost Attractor Networks are introduced as efficient dynamical decoders for closed-loop sequential generation, significantly reducing parameters and latency.

Cybersecurity Hacker News

The Australian Government to Require SMS/MMS Sender ID Registraion

The Australian government is implementing a mandatory registration system for SMS and MMS sender IDs to combat spam and fraud.

AI/ML Hacker News

DeepSeek Introduces Vision

DeepSeek has released a new Vision model, expanding its capabilities to include visual understanding.

Other Hacker News

How to Become a Person After Smartphones Have Rotted Your Brain

An essay exploring the cognitive impact of smartphones and offering strategies to regain focus and mental presence.

AI/ML arXiv cs.AI

EMORSION: Examining the Impact of Audio Parameters on Emotional Responses and Immersion in Film

The EMORSION study examines how manipulating audio parameters like pitch and loudness in films affects audience emotion and immersion.

AI/ML arXiv cs.AI

Towards Multi-Agent-Simulation-Based Community Note Evaluation

MultiCom is a multi-agent framework that uses persona-guided agents to evaluate community notes for fact-checking on social media.

AI/ML arXiv cs.AI

Mitigating Anchoring Bias in LLM-Based Agents for Energy-Efficient 6G Autonomous Networks

A framework for 6G autonomous networks that uses LLM agents and a randomized anchoring strategy to reduce energy consumption and mitigate bias.

AI/ML arXiv cs.AI

Continuous Audio Thinking for Large Audio Language Models

Continuous Audio Thinking (CoAT) provides LALMs with a latent workspace to better preserve and organize acoustic information before generating responses.

Software Engineering arXiv cs.AI

IOAH3: Importance-Driven Adaptive Spatial Partitioning

IOAH3 introduces an adaptive spatial partitioning method using H3 grids and importance scoring to reduce the modifiable areal unit problem in geo-spatial analysis.

AI/ML arXiv cs.AI

Breaking the Solver Bottleneck: Training Task Generators at the Learnable Frontier

PROPEL is a solver-amortized framework that trains task generators to produce RL training tasks at the targeted 'learnable frontier' for agents.