MathArena - BrokenArXiv March (Iterative Self-Correction): leaderboard
A variant of the MathArena BrokenArXiv benchmark (2026): models are asked to prove plausible but false statements derived from recent arXiv papers and may critique and revise their answer up to twice. The score is accuracy in percent on the 56 March 2026 problems after self-correction, where a response succeeds if it flags the statement as false or declines to prove it; higher is better.
Metric: BrokenArXiv score (%) on the 56 problems of the March 2026 BrokenArXiv set (plausible statements from recent arXiv papers perturbed to be false), MathArena's provider-recommended settings, no tools, asked to prove each false statement, with up to two rounds of self-critique and revision after the first attempt (Huang and Yang verifier-and-revision loop, at most three proof attempts); higher is better. Source: arxiv.org. Saturation forecast: Around August 2027. 3 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 34.8 | #99 |
| 2 | Step 3.5 Flash | 27.8 | #526 |
| 3 | MathArena GLM 5.1 (checkpoint unspecified) | 12.5 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=matharena-brokenarxiv-march-iterative-self-correction · How It Works · Data refreshed daily, snapshot 2026-11-21.