PhysCount - Image Formation NAE: leaderboard
PhysCount (arXiv 2610.40530): 4,595 human-written counting questions on 546 real photos with mirrors and glass, where one physical object can appear several times by reflection or transmission. Normalized absolute error on counts by how each appearance is observed: directly, by reflection or through glass (lower is better).
Metric: Image-formation normalized absolute error (%). Source: arxiv.org. Saturation forecast: Around November 2027. 40 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 27B | 26.78 | #192 |
| 2 | Qwen 3.5 397B A17B | 26.88 | #154 |
| 3 | Qwen 3.5 122B A10B | 27.8 | #184 |
| 4 | Qwen 3.5 35B A3B | 28.49 | #268 |
| 5 | Muse Glimmer 30B | 28.69 | #191 |
| 6 | Qwen 3.8 27B | 29.66 | #123 |
| 7 | Qwen 3.5 9B | 30.05 | #391 |
| 8 | Gemini 3.7 Flash | 30.62 | #24 |
| 9 | GPT-6 Astra | 30.99 | #6 |
| 10 | GPT-5.6 Sol | 31.94 | #22 |
| 11 | Qwen 3.5 4B | 32.93 | #514 |
| 12 | Gemma 4 31B (IT) | 33.71 | #178 |
| 13 | Molmo2-8B | 34.35 | |
| 14 | GPT-5.6 Luna | 35.07 | #79 |
| 15 | Gemma 4 26B A4B (IT) | 36.53 | #293 |
No result here: #3 Claude Opus 5.5, #7 GPT-6.1 Sol, #9 Gemini 4 Argon.
Interactive version: theaggregate.ai/benchmark?slug=physcount-image-formation-nae · How It Works · Data refreshed daily, snapshot 2026-11-05.