PRISM-VLM - Multi-Question Robustness: leaderboard
Metric: Per-axis mean score (%; share of items whose 3-5 bundled questions about one image are all answered correctly; over PRISM-VLM's 6,238 items recycled from 15 public VLM benchmarks (five seeds of 100 items per benchmark); GPT-5 (low effort) synthesizes the perturbations and grades the open-ended axes; each model at the lowest reasoning effort its provider exposes, temperature 0). Source: arxiv.org. Saturation forecast: Around 2029. 42 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3 Flash (Minimal) | 56 |
| 2 | GPT-4.1 Mini | 50.7 |
| 3 | Qwen 3.5 9B (Non-reasoning) | 50.5 |
| 4 | GPT-5 Mini (Minimal) | 49.4 |
| 5 | Gemini 2.5 Flash (Non-reasoning) | 48.7 |
| 6 | GPT-5.4 Mini | 48.6 |
| 7 | Qwen 3.5 4B (Non-reasoning) | 48.5 |
| 8 | Gemini 2.0 Flash | 46.8 |
| 9 | Qwen 3 VL 8B Instruct | 44.4 |
| 10 | Gemini 2.0 Flash Lite | 43.7 |
| 11 | Qwen 3 VL 4B Instruct | 41.5 |
| 12 | Gemini 2.5 Flash Lite | 40.8 |
| 13 | Molmo2-8B | 39.7 |
| 14 | Nova 2 Lite | 38.7 |
| 15 | Claude Haiku 4.5 | 37.4 |
Interactive version: theaggregate.ai/benchmark?slug=prism-vlm-multi-question-robustness · How It Works · Data refreshed daily, snapshot 2026-09-26.