AI/ML arXiv cs.AI

Retraction-Free Optimization over the Stiefel Manifold for the LoRA Fine-Tuning

Researchers propose a retraction-free optimization method for LoRA fine-tuning using the Stiefel manifold to accelerate training.

AI/ML arXiv cs.AI

CAST: Game Solvers as Turn-Level Teachers for LLM Agents

CAST is a method that uses game solvers as teachers to provide turn-level credit assignment signals for LLM agents, improving performance in long-horizon tasks.

AI/ML arXiv cs.AI

Balanced Soft mixture-of-expert model for Glaucoma Detection

A balanced soft mixture-of-experts model is proposed to improve the accuracy and efficiency of multi-modal glaucoma detection.

AI/ML arXiv cs.AI

Physics-Informed Neural Operator for Warm-Starting Background-Decomposed and Preconditioned PSFD: Enabling Scalable 3-D EUV Mask Simulation

A physics-informed neural operator is introduced to enable scalable 3-D EUV mask simulation by accelerating background-decomposed PSFD solvers.

Cybersecurity arXiv cs.AI

Specula: Scaling formal specifications for autonomous model checking of system code

Specula is an autonomous agentic system that uses LLMs to generate formal TLA+ specifications for system code to facilitate model checking and bug finding.

AI/ML arXiv cs.AI

Every Time I Hire a Linguist, Inference Costs Go Down: On Linguistic Rules as Effective Prompt Compressors

Researchers explore using deterministic linguistic rules as effective, CPU-side prompt compressors for LLMs to reduce inference costs.

AI/ML arXiv cs.AI

Explainable AI for Chronic Kidney Disease Prediction Using Simulated Federated Learning

A study uses Federated Learning and Explainable AI (XAI) to predict Chronic Kidney Disease with high accuracy and transparency.

Hardware/Chips Hacker News

Balcony Solar: $1.74/watt direct plug-in solar

Discussion on low-cost balcony solar panels offering $1.74 per watt for direct plug-in installation.

AI/ML Hacker News

Claude Is Down

Reports of service outages for the Claude AI assistant.

Cybersecurity TechCrunch

The Hugging Face AI break-in, as told through an increasingly committed bear metaphor

An analysis of a security breach at Hugging Face using a metaphorical approach.

Tech Business/VC VentureBeat

Enterprise AI agents can't talk to each other, can't be trusted with permissions, and can't be audited — 5 startups are already fixing that

An overview of five startups focusing on the infrastructure, orchestration, and security layers needed for enterprise AI agents to communicate and be audited.

AI/ML arXiv cs.AI

Structure-aware Relative Policy Optimization for Ranking

Introduces SRPO, a structure-aware relative policy optimization framework to improve the stability and effectiveness of listwise ranking in RL.

AI/ML arXiv cs.AI

Where Steering Signals Come From: Activation Source Selection in Activation Steering

Research on activation steering in LLMs, identifying that the most effective steering signals originate from execution-boundary states.

AI/ML arXiv cs.AI

Bridging Compute- and Data-Optimal Pretraining

Proposes Compute-Data (CD) scaling laws to bridge the gap between compute-optimal and data-optimal pretraining when high-quality data is limited.

AI/ML arXiv cs.AI

ScaleResfusion: Residual Rectified Flow based on Residual Vector Field

Presents ScaleResfusion, a scalable diffusion framework for real-world image restoration using residual rectified flow and pre-trained models.

AI/ML arXiv cs.AI

CLBench-V: Evaluating Multimodal Context Learning from Grounding to Knowledge Acquisition

Introduces CLBench-V, a benchmark for evaluating multimodal context learning across dimensions like grounding and knowledge acquisition.

Cybersecurity arXiv cs.AI

Hybrid Analysis for Secure MCP Tool Use in LLM Agents

Proposes MTGuard, a hybrid analysis-based defense framework to secure the use of MCP tools in LLM agents against malicious actions.

Other Hacker News

The motion-sickness cure hidden in iPhone settings

A discussion on a specific iPhone setting that can help reduce motion sickness.

Homelab/Self-Hosting Hacker News

Turning a Dumb AC Unit Smart (Without Losing My Security Deposit)

A project about making a standard AC unit smart while adhering to rental constraints.

Software Engineering Hacker News

How to Not Die by a Thousand Cuts. Or, How to Think About Software Quality

A conceptual piece on software quality and how to avoid systemic degradation in software projects.