Technology & AI
Best AI agents fail 50% of visual tool tasks, Apple benchmark shows
A new open benchmark with 500+ tools reveals even frontier AI models can't reliably read an image and act on it, with most failures traced to seeing, not
AI Agents
AI agents explained — autonomous assistants, coding agents and the research on where they work and where they fail.
117 stories
Technology & AI
Lean-QIT: machine-checked proofs for quantum information theory
A new Lean 4 library formalises three quantum coding theorems and builds reusable infrastructure for AI-assisted proof search in quantum information
Technology & AI
Malaika AI framework uses tri-grounded reasoning for malware analysis
Malaika, a multi-agent AI framework posted to arXiv this month, uses three grounding mechanisms to make malware analysis more precise and auditable.
Technology & AI
TrustX ARC rates AI agent risk across 12 dimensions
TrustX ARC, a new arXiv preprint, scores agentic AI risk across 12 dimensions into three governance tiers for risk officers, developers and regulators.
Technology & AI
LLM agent framework blocks hallucinated actions in industrial control
A new arXiv preprint pairs an LLM planner with a forecasting model to guard industrial control systems, recording zero hallucinated actions in attack
Technology & AI
KV-PRM cuts AI agent scoring cost 5,000x via cache reuse
KV-PRM reads the memory AI models already produce during generation, slashing the cost of verifying multi-agent reasoning chains by up to 5,000x.
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
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
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.