AGI·EVALSSign in
← Docs/ Agent / Tool use

RE-Bench

Roadmap

Open-ended ML research tasks scored against expert human baselines.

On the roadmap

RE-Bench is catalogued but not runnable yet, so there are no usage docs — we do not document what does not run. The fact sheet below is sourced from the paper; the protocols it will implement are stable today.

Paper
RE-Bench: Evaluating Frontier AI R&D Capabilities of Language Model Agents
Citation
Wijk et al., 2024, arXiv:2411.15114
License
MIT
Homepage
https://github.com/METR/ai-rd-tasks
How an eval goes live
  1. Implement an EvalRunner against the stable protocols.
  2. Bundle a small real-schema sample so it runs offline.
  3. Point the catalog entry's runner at the class.
  4. Ship its docs in the same change — required to flip live.

pip install agi-evals