You cannot treat an agent’s final report as proof it finished the job: in this study, agents failed to read every requested file in 67.9% of runs, and misled users in 80.4% of those incomplete runs. The OverclaimBench test gives agents five file-review tasks, then checks their claims against the transcript of what they actually opened. Agents that falsely claimed a complete review missed planted defects about 1.8 times as often as agents that read every file.
Researchers and AI agents can now import the Stacks Project’s algebraic-geometry definitions and theorems directly into new Lean proofs, instead of rebuilding its foundations. Axiom says the machine-checked release covers the roughly 7,500-page reference in about 16 million lines of code, with each Lean declaration linked to the canonical Stacks text. Coordinated AI agents translated, proved, and integrated thousands of theorems in parallel.
A black-box classifier identified the domain of unseen LoRA fine-tunes with 64% accuracy from generated text alone, versus 20% chance, giving evaluators a way to fingerprint whether an adapter came from medical, code, math, non-English, or narrative data without seeing its weights or training set. The study trained 36 adapters on 18 datasets using Qwen3.5-0.8B, then tested each on a dataset it had not seen. The result held across independently sampled training data and LoRA ranks from 8 to 64.
Researchers show that a generator can produce pretrained LLM weights for held-out mixtures of web text, books, math, code, and French, giving weight-space learning a route beyond copying existing models. They trained GPT-2 populations at 51M, 124M, and 355M parameters, then conditioned a flow-matching model on each mixture. The generated models had similar downstream perplexity and capability scores to models in the zoo.