ClosureBench - L3 Composition: leaderboard
Metric: Accuracy (%; lenient scoring of one JSON answer per question; L3 compositional queries (three to five steps, 350 instances); held-out split of 1,300 generated graph questions over graph sizes 4-10, temperature 0; direct mode: answer immediately). Source: arxiv.org. Saturation forecast: Around December 2026. 9 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | O3 | 82 |
| 2 | Gemini 2.5 Flash | 73.4 |
| 3 | Claude Sonnet 4 | 64 |
| 4 | GPT-4.1 | 36.9 |
| 5 | GPT-4.1 Mini | 34.9 |
| 6 | Qwen 3.5 9B | 26.3 |
| 7 | Qwen 3.5 2B | 23.1 |
| 8 | Qwen 3 4B | 22.3 |
| 9 | Qwen 2.5 1.5B Instruct | 15.4 |
Interactive version: theaggregate.ai/benchmark?slug=closurebench-l3-composition · How It Works · Data refreshed daily, snapshot 2026-09-26.