PRISM-VLM - Visual Discrimination: leaderboard
The visual-discrimination axis of PRISM-VLM (NAVER, 2026): the model picks the accurate caption for an image from a pair in which the other caption changes one detail, such as a color, count or position. Accuracy in percent with chance at 50; higher is better.
Metric: Per-axis mean score (%; forced A/B choice between the gold caption and a one-detail minimal edit (chance 50%); 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: Estimated already saturated. 42 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash (Minimal) | 93.7 | #159 (Gemini 3 Flash) |
| 2 | Gemini 3.1 Flash Lite (Minimal) | 92.6 | #363 (Gemini 3.1 Flash Lite) |
| 3 | Gemini 2.5 Flash (Non-reasoning) | 91.5 | #452 (Gemini 2.5 Flash) |
| 4 | Gemini 2.0 Flash | 90 | #702 |
| 5 | Gemini 2.0 Flash Lite | 87.6 | #1032 |
| 6 | Gemini 2.5 Flash Lite | 87.5 | #958 |
| 7 | GPT-4.1 Mini | 86.7 | #775 |
| 8 | Molmo2-8B | 86.4 | #1488 |
| 9 | GPT-5 Mini (Minimal) | 86 | #349 (GPT-5 Mini) |
| 10 | Qwen 3.5 9B (Non-reasoning) | 85.9 | #803 (Qwen 3.5 9B) |
| 11 | Qwen 3 VL 8B Instruct | 85.8 | #982 |
| 12 | Gemma 4 E4B | 84 | #2217 |
| 13 | Claude Haiku 4.5 | 83.9 | #568 |
| 14 | Qwen 3 VL 4B Instruct | 83 | #1281 |
| 15 | GPT-5.4 Mini | 82.7 | #389 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=prism-vlm-visual-discrimination · How It Works · Data refreshed daily, snapshot 2026-11-12.