WildTableBench - Value Modification: leaderboard

A WildTableBench (2026) subtype on recomputing an aggregate after specified cell values change, asked about real-world table images from the web. Accuracy in percent, free-form answers judged against the reference by GPT-5.2; higher is better.

Metric: Accuracy (%) on the Value Modification subtype (C4-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 February 2027. 21 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-5.2High66.7#185
2Gemini 3 Pro61.9#145
3Gemini 3 Flash57.1#175
4Kimi K2.5Thinking52.4#266
5Qwen 3 VL 235B A22B (Thinking)42.9#459
6Claude Sonnet 4.6High42.9#173
7Seed 2.0 ProHigh42.9#222
8Qwen 3 VL 32B Instruct33.3#606
9Claude Opus 4.6High33.3#113
10GPT-5 MiniHigh28.6#364
11Qwen 3 VL 235B A22B Instruct23.8#532
12GLM-4.6V23.8#710
13O3High9.5#237
14Qwen 3 VL 4B Instruct4.8#1332
15Qwen 3 VL 32B (Thinking)4.8#573

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=wildtablebench-value-modification · How It Works · Data refreshed daily, snapshot 2026-11-21.