WildTableBench - Multi-step Conditional: leaderboard
Metric: Accuracy (%) on the Multi-step Conditional subtype (C2-M), 928 human-verified questions over 402 real-world table images, free-form answers judged by GPT-5.2 against the reference answer, high reasoning effort where configurable; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 21 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3 Pro | 66.4 |
| 2 | Gemini 3 Flash | 51 |
| 3 | GPT-5.2 (High) | 49 |
| 4 | Kimi K2.5 (Thinking) | 45.5 |
| 5 | Claude Opus 4.6 (High) | 45.5 |
| 6 | Seed 2.0 Pro (High) | 42 |
| 7 | Claude Sonnet 4.6 (High) | 34.3 |
| 8 | Qwen 3 VL 235B A22B (Thinking) | 32.2 |
| 9 | Qwen 3 VL 235B A22B Instruct | 25.2 |
| 10 | Qwen 3 VL 32B Instruct | 23.8 |
| 11 | GLM-4.6V | 23.8 |
| 12 | Qwen 3 VL 32B (Thinking) | 23.8 |
| 13 | GPT-5 Mini (High) | 21.7 |
| 14 | O3 (High) | 14 |
| 15 | Qwen 3 VL 8B (Thinking) | 8.4 |
Interactive version: theaggregate.ai/benchmark?slug=wildtablebench-multi-step-conditional · How It Works · Data refreshed daily, snapshot 2026-10-07.