FaithformBench - Omni-MATH - Unfaithfulness Lower Bound: leaderboard
FaithformBench (arXiv 2608.10916v1, PDF Figures 3 and 4): faithfulness when autoformalising ProcessBench reasoning steps into Lean, checked by proving each formalisation or its negation for original and invalidated steps. Share of original and perturbed Omni-MATH steps with a failure: a valid step refuted, an invalid step proved (silent correction) or no usable formalisation; a lower bound on unfaithfulness, percent. Lower is better.
Metric: Unfaithfulness lower bound (%; half FNR + half FPR + AF failure rate). Source: arxiv.org. Saturation forecast: Around February 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 14 | #51 |
| 2 | Gemini 3.1 Pro (Preview) | 18 | #64 |
| 3 | GPT-5.2 | 26 | #115 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=faithformbench-omni-math-unfaithfulness-lower-bound · How It Works · Data refreshed daily, snapshot 2026-11-05.