See-or-Guess VQA - Vision-Grounded: leaderboard
See-or-Guess VQA (2026) measures how much vision-language models answer from textual priors by rewriting questions about images from 21 public benchmarks into variants that draw less and less on the source question. This board is the vision-grounded variant, with questions written from the image alone, scored as accuracy in percent judged by Qwen3-14B; higher is better.
Metric: Vision-Grounded variant (questions Gemini-2.5-Pro writes from the image alone, no source text): accuracy (%) with the image over 540 images from 21 public benchmarks in six reasoning categories; Qwen3-14B judges answer equivalence (yes/no), lenient extraction of answers after a missing </think> tag. Source: arxiv.org. Saturation forecast: Around October 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 9B | 42 | #803 |
| 2 | Qwen 3.5 397B A17B | 38.4 | #246 |
| 3 | Gemini 2.5 Pro | 35.9 | #266 |
| 4 | GPT-5 Mini | 33.3 | #348 |
| 5 | Claude Sonnet 4.6 | 31.6 | #161 |
| 6 | Gemini 3.1 Flash Lite | 27.3 | #364 |
| 7 | Phi-4 Multimodal Instruct | 10 | #3882 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=see-or-guess-vqa-vision-grounded · How It Works · Data refreshed daily, snapshot 2026-11-13.