ClosureBench: leaderboard
ClosureBench (Goria, 2026) generates graph-relational reasoning questions over small directed graphs and scores single JSON answers against reference programs, with lenient tolerances for documented ambiguities. In direct mode the model answers immediately; this board covers all 1,300 held-out questions at three compositional levels. Accuracy in percent; higher is better.
Metric: Accuracy (%; lenient scoring of one JSON answer per question; all three compositional levels, instance-weighted; 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: Estimated already saturated. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O3 | 92.8 | #224 |
| 2 | Gemini 2.5 Flash | 87.4 | #452 |
| 3 | Claude Sonnet 4 | 81.7 | #397 |
| 4 | GPT-4.1 | 52.4 | #492 |
| 5 | GPT-4.1 Mini | 49.2 | #775 |
| 6 | Qwen 3.5 9B | 43.3 | #803 |
| 7 | Qwen 3 4B | 39.5 | #2297 |
| 8 | Qwen 3.5 2B | 33 | #4849 |
| 9 | Qwen 2.5 1.5B Instruct | 25.8 | #7518 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=closurebench · How It Works · Data refreshed daily, snapshot 2026-11-12.