HoloCount - Visual Distractor Illusion: leaderboard
Metric: Accuracy (%; exact-match accuracy of the single integer count a multimodal model returns zero-shot for an image and a counting question, with a strict integer-only system prompt; counting real objects among look-alike distractors such as plastic fruit (200 questions); higher is better). Source: arxiv.org. Saturation forecast: Around March 2027. 30 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Qwen 3.5 397B A17B | 80 |
| 2 | Kimi K2.6 | 78.5 |
| 3 | Qwen 3.5 122B A10B | 78 |
| 4 | Qwen 3.5 35B A3B | 77.5 |
| 5 | Qwen 3.5 9B | 76.5 |
| 6 | GPT-5.5 | 76 |
| 7 | Kimi K2.5 | 73.5 |
| 8 | Claude Opus 4.8 | 73.5 |
| 9 | GPT-4.1 | 73 |
| 10 | Qwen 3.5 27B | 73 |
| 11 | Gemini 2.5 Pro | 71 |
| 12 | Qwen 3 VL 32B Instruct | 71 |
| 13 | Claude Opus 4.7 | 70.5 |
| 14 | Qwen 3.5 4B | 70.5 |
| 15 | MiniMax-M3 | 69 |
Interactive version: theaggregate.ai/benchmark?slug=holocount-visual-distractor-illusion · How It Works · Data refreshed daily, snapshot 2026-09-29.