An open framework combines independent peer review, a specialised wiki and Lean verification. The authors report high scores on several scientific tests from 2026.
Humanize Olympiad Agents organizes the resolution of difficult problems around several sessions. A critical review of the candidates, a wiki provides targeted knowledge and a Lean checker checks the formal proofs against the reference statements.
With GPT-5.6 Sol, the authors report the resolution of 672 PutnamBench problems and 251 of 304 problems in Lean-Eval. They also evaluate the framework with Kimi-K3. The results illustrate the contribution of the feedback loop and formal control in these tests; they do not describe the participation of a system in official competitions.
Preprint filed on arXiv on October 9, 2026; results are those reported by the authors and remain to be independently confirmed.
Read the full article on the original source
Go to the source →