ClosureBench (Chain-of-Thought) - L1 Atomic: leaderboard

Metric: Accuracy (%; lenient scoring of one JSON answer per question; L1 atomic queries (one reasoning step, 550 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: Estimated already saturated. 4 models tracked.

Top models

#ModelScore
1Gemini 2.5 Flash91.8
2GPT-4.191.5
3GPT-4.1 Mini91.1
4Claude Sonnet 490.5

Interactive version: theaggregate.ai/benchmark?slug=closurebench-chain-of-thought-l1-atomic · How It Works · Data refreshed daily, snapshot 2026-09-26.