HoloCount - Linguistic Prior Conflict: 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; images that contradict a strong language prior, such as a lobster with other than two claws (163 questions); higher is better). Source: arxiv.org. Saturation forecast: Around November 2027. 30 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 73.6 |
| 2 | Gemini 3 Flash (Preview) | 69.9 |
| 3 | Claude Opus 4.8 | 52.1 |
| 4 | Qwen 3.5 27B | 52.1 |
| 5 | Qwen 3.5 397B A17B | 50.3 |
| 6 | Claude Opus 4.7 | 49.1 |
| 7 | GPT-5.5 | 48.5 |
| 8 | Kimi K2.6 | 48.5 |
| 9 | Qwen 3.5 122B A10B | 47.9 |
| 10 | GPT-5.4 | 46.6 |
| 11 | GPT-4.1 | 46.6 |
| 12 | Qwen 3.5 35B A3B | 46.6 |
| 13 | Kimi K2.5 | 44.8 |
| 14 | Gemini 2.5 Pro | 44.2 |
| 15 | Qwen 3.5 4B | 44.2 |
Interactive version: theaggregate.ai/benchmark?slug=holocount-linguistic-prior-conflict · How It Works · Data refreshed daily, snapshot 2026-09-29.