AI/ML arXiv cs.AI

SLAC: Safe and Efficient Real-Robot Reinforcement Learning via Unsupervised Simulation Pre-Training

SLAC is introduced as a method for real-robot reinforcement learning that uses unsupervised simulation pre-training to create a safe, task-agnostic latent action space.

AI/ML arXiv cs.AI

A Bit of Freedom Goes a Long Way: Classical and Quantum Algorithms for Reinforcement Learning under a Generative Model

The paper proposes new classical and quantum algorithms for reinforcement learning that break the classical regret bound barrier using a generative model simulator.

AI/ML arXiv cs.AI

Acoustic Imaging for UAV Detection: Dense Beamformed Energy Maps and U-Net SELD

A new U-net based approach for 360-degree acoustic source localization is presented, treating UAV detection as a spherical semantic segmentation task.

AI/ML arXiv cs.AI

Unsupervised Deep Learning for Inverse Problems in Computed Tomography

An unsupervised deep learning framework is proposed for Computed Tomography (CT) reconstruction, achieving significant speed-ups over per-image optimization methods.

Cybersecurity arXiv cs.AI

Latent Fusion Jailbreak: Blending Harmful and Harmless Representations to Elicit Unsafe LLM Outputs

Latent Fusion Jailbreak (LFJ) is a white-box attack that manipulates internal LLM representations to elicit unsafe outputs with high success rates.

Cybersecurity arXiv cs.AI

Poison to Detect: Detection of Targeted Overfitting in Federated Learning

The paper proposes three detection techniques to identify targeted overfitting attacks by dishonest orchestrators in Federated Learning systems.

AI/ML arXiv cs.AI

A Systematic Study of Large Language Models for Task and Motion Planning With PDDLStream

A systematic study evaluates the integration of LLMs with PDDLStream for task and motion planning in robotics, finding LLM-based planners often underperform compared to engineered systems.

Other Hacker News

What does the Riemann zeta function have to do with the distribution of primes?

A discussion on the relationship between the Riemann zeta function and the distribution of prime numbers.

Tech Business/VC Hacker News

Airbus Takes Flight from AWS

Airbus is moving its infrastructure away from Amazon Web Services (AWS).

AI/ML arXiv cs.AI

Towards a General Intelligence and Interface for Wearable Health Data

Researchers propose a foundation model for wearable health data pretrained on one trillion minutes of sensor signals from five million participants to provide personalized health insights.

AI/ML arXiv cs.AI

Workflow-GYM: Towards Long-Horizon Evaluation of Computer-use Agentic tasks in Real-World Professional Fields

Workflow-GYM is introduced as a benchmark for evaluating AI agents' ability to perform long-horizon GUI tasks in professional software environments.

AI/ML arXiv cs.AI

Agents-K1: Towards Agent-native Knowledge Orchestration

Agents-K1 is an end-to-end knowledge orchestration pipeline that converts scientific papers into agent-native knowledge graphs, including the release of the Scholar-KG dataset.

AI/ML arXiv cs.AI

Perception-Aligned AI Outputs: End-to-End Visual Prediction for Uncertainty Communication in Clinical Decision-Making

VL4ML is a human-centered explainability framework that uses visual representations instead of numeric outputs to communicate AI model uncertainty in clinical settings.

AI/ML arXiv cs.AI

Why do CNNs excel at feature extraction? A mathematical explanation

A mathematical study providing a theoretical explanation for why Convolutional Neural Networks (CNNs) are effective at feature extraction for image classification.

AI/ML arXiv cs.AI

CTC: The Composite Task Challenge for Cooperative Multi-Agent Reinforcement Learning

The Composite Tasks Challenge (CTC) is proposed as a benchmark to evaluate division of labor and cooperation in multi-agent reinforcement learning.

AI/ML arXiv cs.AI

MAnchors: Memorization-Based Acceleration of Anchors via Rule Reuse and Transformation

MAnchors proposes a memorization-based framework to accelerate the Anchors local model-agnostic explanation technique without losing fidelity.

AI/ML arXiv cs.AI

AuditVotes: Elevating Provable Defense for GNNs with Efficient Augmentation and Conditional Smoothing

AuditVotes is a framework for Graph Neural Networks (GNNs) that improves both clean accuracy and certified robustness against adaptive attacks.

AI/ML The Verge

China delivers a one-two punch to America’s AI dominance 

Chinese AI companies Moonshot and Alibaba have released new models, including Kimi K3, that claim to rival US models at lower costs.

AI/ML arXiv cs.AI

When Does Muon Help Agentic Reinforcement Learning?

Research indicates that the Muon optimizer can significantly improve success rates in agentic reinforcement learning, particularly in sparse-reward environments.

Cybersecurity arXiv cs.AI

Evaluating Open-Weight LLMs for Generating Structured Threat Information for Autonomous Vehicle Vulnerabilities

Researchers evaluated open-weight LLMs for generating structured threat information (STIX) to automate vulnerability analysis for autonomous vehicles.