MIRA-Math: leaderboard
MIRA-Math (2026) gives a model a math problem with exactly one necessary fact withheld; the model must ask a fixed LLM responder for that fact within a small request budget and then compute the exact answer. Zero-shot final-answer accuracy in percent, higher is better.
Metric: Final-answer accuracy (%; exact verifier match; zero-shot; 2,310 generated instances from 22 typed families, one hidden atomic fact per instance, request budget 3/6/9 by difficulty, fixed gpt-4o-mini responder, temperature 0, single pass; fraction x100). Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemma 4 31B (IT) | 73.9 | #342 |
| 2 | GPT-5.1 | 69 | #229 |
| 3 | Grok 4.20 | 50.3 | #297 |
| 4 | Gemini 2.5 Flash | 49 | #453 |
| 5 | GPT-4o Mini | 24.8 | #1623 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mira-math · How It Works · Data refreshed daily, snapshot 2026-11-13.