AgentVQA
Multimodal · 2026-10-22
1,000 agentic visual-understanding questions from Princeton University and the University of Maryland, 200 in each of five domains: web agents, egocentric videos, games, spatial planning and robotics. Questions are manually verified four-option multiple-choice items converted from 23 existing datasets, so offline answers stand in for online agent actions such as choosing the next GUI click or predicting a robot's next step. Models answer with a single letter from the visual input (images, screenshots or video) and any action history; chance is 25%. The score is the percentage answered correctly over all 1,000.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 3.7 Flash | 58.5 |
| Gemini 3.1 Pro Preview | 55.7 |
| GPT-5.6 Sol | 55.0 |
| Qwen3.7 Flash | 55.0 |
| Qwen3.8 27B | 50.9 |
| Qwen3.7-Plus (source unspecified) | 48.6 |
| Kimi K3 | 48.3 |
| GPT-5.6 Luna | 47.9 |
| Qwen3.5 35B A3B | 47.9 |
| Qwen3.5 397B A17B | 47.8 |
| Qwen3.5-27B | 46.7 |
| Qwen3.5 122B A10B | 46.0 |
| Opus 4.8 | 44.0 |
| Sonnet 5 | 43.4 |
| GLM-4.5V | 42.9 |
| GLM-4.6V | 41.8 |
| GLM-4.1V-9B-Thinking | 35.4 |
| GPT-4o (source unspecified) | 31.4 |
| Llama 4 Maverick Instruct | 28.0 |