RefGlitch-Bench: leaderboard
RefGlitch-Bench (2026) tests whether vision-language models spot visual glitches in video game frames when shown a clean reference frame of the same scene. This board is accuracy from 0 to 1 for binary glitch detection on a balanced set of glitchy and glitch-free frames, each shown with a manually selected clean reference frame; higher is better.
Metric: Accuracy (fraction) of binary gameplay-glitch detection on 1,000 balanced test frames (500 glitchy, 500 glitch-free) with a manually selected clean reference frame (Ref setting); higher is better. Source: arxiv.org. Saturation forecast: Around January 2028. 5 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 0.77 | #178 |
| 2 | Qwen 3 VL 8B | 0.73 | #917 |
| 3 | Ministral 3 8B | 0.68 | #2131 |
| 4 | Gemini 3 Flash (Preview) | 0.63 | #152 |
| 5 | Gemma 3 12B | 0.63 | #2059 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=refglitch-bench · How It Works · Data refreshed daily, snapshot 2026-11-21.