VistaQA - Answer Accuracy: leaderboard
VistaQA (2026) asks models for a free-form answer to a visual question together with segmentation masks of the supporting evidence, over 1,157 samples in six task types and six domains, including questions about absent objects. This board is the text answer accuracy alone, judged by an LLM; percent, higher is better.
Metric: Overall text answer accuracy (%), on the 1,157 VistaQA samples (six task types, six visual domains, 314 hallucination samples whose queried entity is absent), zero-shot with standardized output-format instructions; the answer is judged correct or not by Qwen 2.5-14B and the evidence masks are matched to the reference masks by Hungarian IoU; higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 13 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | SAM3 + Gemini 3 (VistaQA checkpoint unspecified) | 62.32 | ||
| 2 | SAM3 + GPT-5.4 | Thinking | 61.02 | |
| 3 | SAM3 + Qwen3-VL-32B-Instruct | 57.65 | ||
| 4 | SAM3 + Qwen3-VL-4B-Instruct | 53.15 | ||
| 5 | R-Sa2VA-Qwen3VL-4B-RL | 36.3 | ||
| 6 | Sa2VA-8B | 29.04 | ||
| 7 | UniPixel-7B | 21.26 | ||
| 8 | TreeVGR-7B | 16.08 | ||
| 9 | LISA-7B | 7.26 | ||
| 10 | SESAME-7B | 3.2 | ||
| 11 | GLaMM-7B | 1.38 | ||
| 12 | UGround-7B | 0.95 | ||
| 13 | LaSagnA-7B | 0.43 |
Interactive version: theaggregate.ai/benchmark?slug=vistaqa-answer-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-21.