AI Research

Study from Google Research, MIT and Harvard: AI models spot flaws but often leave them out of reports. One sentence in the prompt helps. How to use the finding.

Claude reconstructs the origins of historical NIST seeds. What the finding explains about digital signatures, and why it does not break encryption.

TheoPhysEval reports strong AI results on theoretical physics calculations. What 49 solved problems mean, and how researchers can reproduce the evaluation.

AlphaProof Nexus reaches Science. How Lean makes AI proofs checkable, what the results establish, and who can use the published proof files.

Since October, the preprint server arXiv allows only two submissions per person per month. What is behind the emergency brake and what researchers need to know now.

Epoch AI asked agents to develop new training methods. Its test reveals why verified research progress can fall short of persuasive final reports.

Vals AI had GPT-6 Sol and Opus 5.5 code alone and in teams: up to five times the cost, barely better apps. Here is when subagents still make sense.

A study finds nearly 13,000 typical AI phrases, but every model version has its own. What the list reveals about Claude, GPT and Gemini, and where it fails.

An astrophysicist used AI agents in Claude Science to fill the gap in the ultraviolet sky map. How the unmeasured third was made and where the human stepped in.

OpenAI has released 722 math manuscripts from its internal model, 162 of them machine-checked. What is inside, what is missing and why mathematicians are divided.