OrdinalBench (3D Single-Loop) - Stepwise Trace Accuracy: leaderboard
Metric: Stepwise trace accuracy (%): share of steps of the model's counting trace that match the ground-truth traversal, averaged over 9,000 questions on rendered 3D single-loop scenes; zero-shot at temperature 0, the model returns the target and its step-by-step trace as JSON; GPT-5 and Gemini 2.5 Flash Lite were scored on stratified subsets of the public test split; higher is better. Source: arxiv.org. Saturation forecast: Around 2035. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2.5 VL 72B Instruct | 33.25 | #364 |
| 2 | Qwen 2.5 VL 32B Instruct | 31.93 | #443 |
| 3 | Qwen 2.5 VL 7B Instruct | 12.1 | #643 |
| 4 | GPT-5 | 12.05 | #91 |
| 5 | Gemini 2.5 Flash Lite | 9.36 | #413 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=ordinalbench-3d-single-loop-stepwise-trace-accuracy · How It Works · Data refreshed daily, snapshot 2026-10-11.