ClosureBench (Chain-of-Thought) - 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; chain-of-thought mode: reason step by step, then answer). Source: arxiv.org. Saturation forecast: Around December 2026. 4 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-4.1 Mini | 74.9 |
| 2 | GPT-4.1 | 74.6 |
| 3 | Claude Sonnet 4 | 65.1 |
| 4 | Gemini 2.5 Flash | 63.1 |
Interactive version: theaggregate.ai/benchmark?slug=closurebench-chain-of-thought-l3-composition · How It Works · Data refreshed daily, snapshot 2026-09-26.