Technology & AI
LLM agents fail 70% of skill safety tests in SLBench study
A new 86-case benchmark finds Codex and Claude Code agents violate logical constraints in skill files up to 70% of the time, causing privacy leaks and
AI Models
166 stories
Technology & AI
Agora preprint: auction-based routing for LLM agents
New arXiv preprint proposes auction-based routing for LLM agents, sending reasoning steps to the most capable solver rather than the most overconfident.
Technology & AI
Neural network backdoors evade detection even with full weight access
A preprint shows backdoors in feedforward neural networks evade every statistical test, even with full access to all weights, breaking model trust.
Technology & AI
Visual pretraining outperforms text-only for language AI
Training AI models directly on visual documents consistently outperforms text-only pretraining, challenging a core assumption in how foundation models
Technology & AI
New protocol forces AI scientist agents to show their work
A new arXiv paper proposes the Hypothesis Evolution Protocol, making AI agents' scientific reasoning explicit and auditable instead of buried in logs.
Technology & AI
4DR360 radar-camera AI gives self-driving 360° scene awareness
New arXiv preprint 4DR360 treats 3D scene occupancy as a persistent state, reshaping radar-camera fusion for autonomous driving.
Technology & AI
LLM chip design survey maps path to autonomous EDA agents
A new arXiv survey maps how LLMs could move front-end chip design from isolated tasks to autonomous agents, but offers no benchmarks to prove it works.
Technology & AI
ReContext fixes LLM long-context blind spots without retraining
A new arXiv preprint proposes a training-free method that helps language models actually use evidence already sitting inside their 128K-token context
Technology & AI
LLM-as-a-Verifier hits 86.5% on Terminal-Bench V2
LLM-as-a-Verifier treats checking AI answers as a scaling axis, hitting 86.5% on Terminal-Bench V2 and 78.2% on SWE-Bench Verified without extra training.
Technology & AI
VAORA aligns vision-language model reasoning with physical actions
VAORA, a new reward design on arXiv, targets hallucinated reasoning and reasoning-action misalignment in vision-language models on physical tasks.