PhyX-AF - MathVerse Function (Image): leaderboard
Multimodal autoformalization into Lean 4 (PhyX-AF; Xiong et al., 2026): 10 MathVerse function-graph problems whose diagram is given as an image and is needed to state the problem. The model runs as the backbone of the MMFormalizer recursive-grounding framework, and two expert annotators judge whether the resulting formal statement is valid; accuracy in percent, higher is better.
Metric: Human-verified autoformalization accuracy (%; 10 MathVerse function-graph problems with the diagram as image input, Lean 4 output judged valid by two expert annotators). Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 60 | #262 |
| 2 | Gemini 3 Pro | 40 | #135 |
| 3 | GPT-5 | 30 | #163 |
| 4 | Qwen 3 VL 235B A22B Instruct | 10 | #521 |
| 5 | Qwen 2.5 VL 72B Instruct | 0 | #845 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=phyx-af-mathverse-function-image · How It Works · Data refreshed daily, snapshot 2026-11-12.