OCR-MetaReasoning - Meta-Abductive: leaderboard
OCR-MetaReasoning answer score on abductive meta-reasoning items over text-rich images (arXiv 2608.30678v1, Table 2); MRMS averages the three reasoning types. Percentage (100 times the normalized score), higher is better; reasoning modes follow the paper's Table 24.
Metric: Answer Score (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) (Thinking) | 94.1 | #64 (Gemini 3.1 Pro (Preview)) |
| 2 | GPT-5.4 (Medium) | 90.8 | #83 (GPT-5.4) |
| 3 | Claude Sonnet 4.6 (Thinking) | 87.9 | #94 (Claude Sonnet 4.6) |
| 4 | GPT-5.4 Mini (Medium) | 87.1 | #228 (GPT-5.4 Mini) |
| 5 | GPT-5.4 (Non-reasoning) | 85.5 | #83 (GPT-5.4) |
| 6 | Kimi K2.5 (Non-reasoning) | 82.5 | #152 (Kimi K2.5) |
| 7 | Qwen 3 VL 235B A22B (Thinking) | 82.1 | #236 |
| 8 | Grok 4.20 (Non-reasoning) | 78 | #179 (Grok 4.20) |
| 9 | Qwen 3 VL 235B A22B Instruct | 77.7 | #280 |
| 10 | GLM-4.6V (Reasoning) | 76.3 | #330 (GLM-4.6V) |
| 11 | Qwen 3 VL 8B (Thinking) | 73.6 | #423 |
| 12 | Qwen 3 VL 30B A3B (Thinking) | 73.3 | #348 |
| 13 | GPT-5.4 Mini (Non-reasoning) | 70.1 | #228 (GPT-5.4 Mini) |
| 14 | Qwen 3 VL 30B A3B Instruct | 63.2 | #390 |
| 15 | Qwen 3 VL 8B Instruct | 57.6 | #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-meta-abductive · How It Works · Data refreshed daily, snapshot 2026-11-05.