LogicGraph: leaderboard
Metric: Success rate (%): share of problems with at least one fully valid proof path (LogicGraph: 900 natural-language first-order logic problems generated from symbolic proof DAGs with 2 to 19 valid minimal derivation paths (300 each with 2-4, 5-7 and 8 or more paths), every model prompted to write as many independent proofs as possible; each step is auto-formalized by DeepSeek-V3.2-Exp and verified with Prover9; mean of the three tiers); higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro (Preview) | 96.11 | #64 |
| 2 | Qwen 3 235B A22B (Thinking) | 91.22 | #304 (Qwen 3 235B A22B) |
| 3 | DeepSeek V3.2 Exp (Thinking) | 88 | #227 (DeepSeek V3.2 Exp) |
| 4 | Gemini 2.5 Pro | 87.78 | #145 |
| 5 | Gemini 2.5 Flash | 85.56 | #237 |
| 6 | Kimi K2 (Thinking) | 72.22 | #236 (Kimi K2) |
| 7 | Claude Sonnet 4.5 (Thinking) | 61.67 | #138 (Claude Sonnet 4.5) |
| 8 | DeepSeek V3.2 Exp (Non-reasoning) | 61 | #227 (DeepSeek V3.2 Exp) |
| 9 | O3 | 58.33 | #121 |
| 10 | O4 Mini | 57.78 | #172 |
| 11 | Claude Sonnet 4.5 | 53.11 | #138 |
| 12 | QwQ-32B | 50.89 | #410 |
| 13 | Qwen 3 235B A22B (Non-reasoning) | 29.56 | #304 (Qwen 3 235B A22B) |
| 14 | GPT-5.1 | 21.56 | #131 |
| 15 | GLM-4.6 | 21.11 | #246 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=logicgraph · How It Works · Data refreshed daily, snapshot 2026-10-11.