See-or-Guess VQA - Multi-Signal: leaderboard
The multi-signal variant of See-or-Guess VQA (2026), a probe of how much vision-language models rely on textual priors: the source question fused with sub-questions elicited turn by turn, scored as accuracy in percent with the image, judged by Qwen3-14B. Higher is better.
Metric: Multi-Signal variant (source question fused with sub-questions elicited turn by turn in a chat): 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 March 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 9B | 58.7 | #803 |
| 2 | Gemini 2.5 Pro | 58.5 | #266 |
| 3 | Qwen 3.5 397B A17B | 54.8 | #246 |
| 4 | Claude Sonnet 4.6 | 54.3 | #161 |
| 5 | GPT-5 Mini | 49.4 | #348 |
| 6 | Gemini 3.1 Flash Lite | 43.7 | #364 |
| 7 | Phi-4 Multimodal Instruct | 29.1 | #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-multi-signal · How It Works · Data refreshed daily, snapshot 2026-11-13.