See-or-Guess VQA - Subquestion-Guided: leaderboard
The subquestion-guided variant of See-or-Guess VQA (2026), a probe of how much vision-language models rely on textual priors: questions built from single-pass perceptual sub-questions, scored as accuracy in percent with the image, judged by Qwen3-14B. Higher is better.
Metric: Subquestion-Guided variant (built from single-pass perceptual sub-questions): 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 February 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 9B | 56.8 | #803 |
| 2 | Qwen 3.5 397B A17B | 54.6 | #246 |
| 3 | Gemini 2.5 Pro | 53.3 | #266 |
| 4 | GPT-5 Mini | 49.4 | #348 |
| 5 | Claude Sonnet 4.6 | 47.8 | #161 |
| 6 | Gemini 3.1 Flash Lite | 41.8 | #364 |
| 7 | Phi-4 Multimodal Instruct | 22.6 | #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-subquestion-guided · How It Works · Data refreshed daily, snapshot 2026-11-13.