OCR-MetaReasoning: leaderboard
OCR-MetaReasoning tests meta-reasoning by multimodal models on text-rich images (arXiv 2608.30678v1). MRMS, the Meta-Reasoning Macro Score, is the macro average over deductive, inductive and abductive items as a percentage, higher is better, from the paper's Table 2; reasoning modes follow the paper's Table 24.
Metric: Meta-Reasoning Macro Score (%). Source: arxiv.org. Saturation forecast: Around February 2027. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) (Thinking) | 89.3 | #64 (Gemini 3.1 Pro (Preview)) |
| 2 | GPT-5.4 (Medium) | 87.7 | #83 (GPT-5.4) |
| 3 | Claude Sonnet 4.6 (Thinking) | 84.3 | #94 (Claude Sonnet 4.6) |
| 4 | GPT-5.4 Mini (Medium) | 84 | #228 (GPT-5.4 Mini) |
| 5 | GPT-5.4 (Non-reasoning) | 83 | #83 (GPT-5.4) |
| 6 | Qwen 3 VL 235B A22B (Thinking) | 81.5 | #236 |
| 7 | Kimi K2.5 (Non-reasoning) | 79.6 | #152 (Kimi K2.5) |
| 8 | Qwen 3 VL 235B A22B Instruct | 77 | #280 |
| 9 | GLM-4.6V (Reasoning) | 75.9 | #330 (GLM-4.6V) |
| 10 | Grok 4.20 (Non-reasoning) | 75.2 | #179 (Grok 4.20) |
| 11 | Qwen 3 VL 30B A3B (Thinking) | 74.7 | #348 |
| 12 | Qwen 3 VL 8B (Thinking) | 74.5 | #423 |
| 13 | GPT-5.4 Mini (Non-reasoning) | 72 | #228 (GPT-5.4 Mini) |
| 14 | Qwen 3 VL 30B A3B Instruct | 65 | #390 |
| 15 | Qwen 3 VL 8B Instruct | 59 | #419 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=ocr-metareasoning · How It Works · Data refreshed daily, snapshot 2026-11-05.