TableVista - Noise: leaderboard

Accuracy under the Noise rendering (speckle, downsampling, tilt and watermarks) of all questions in TableVista (2026), a multimodal table reasoning benchmark in which models answer questions about rendered table images directly with thinking disabled, scored by exact match backed by a semantic-equivalence check. Percent; higher is better.

Metric: Accuracy (%) on the 3,000 TableVista questions under the Noise rendering (Gaussian speckle, downsampling, tilt and watermarks on the Web tables), direct-output prompt with thinking disabled (reasoning effort none for the GPT models), exact match with a GPT-5-mini semantic-equivalence check on exact-match failures; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 29 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-5.4Non-reasoning70.8#135
2Gemma 4 31B (IT)54.3#350
3Llama 4 Maverick Instruct FP853.5#1023
4Qwen 2.5 VL 72B Instruct51.1#841
5GPT-5.4 MiniNon-reasoning49.5#406
6Qwen 2.5 VL 32B Instruct49.3#1062
7Qwen 3.5 27BNon-reasoning49.2#375
8Qwen 3.5 122B A10BNon-reasoning49#346
9Llama 4 Scout Instruct47.5#1579
10InternVL3.5-30B-A3B45.6#1267
11InternVL3.5-14B45.3#1504
12Gemma 4 26B A4B (IT)43.2#556
13Qwen 3 VL 30B A3B Instruct43#895
14Qwen 2.5 VL 7B Instruct42.5#1902
15Qwen 3.6 35B A3BNon-reasoning41.6#481

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

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