PostcondBench - Correctness: leaderboard
Corr@1 in percent: the expected share of methods whose LLM-generated formal postconditions pass every test, for 420 documented methods from 121 Python and Java repositories (PostcondBench, 2026), one sampled postcondition set per method, averaged over code-only, comment-only and code-plus-comment inputs. Higher is better.
Metric: Corr@1 times 100 (%): expected share of methods whose generated postcondition set passes every test, one sampled postcondition set per method (k = 1, five samples drawn), averaged over the Python and Java splits (210 repository methods each) and three input settings (code only, comments only, code plus comments); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | Claude Sonnet 4.5 | Thinking | 62.9 | #261 |
| 2 | GPT-5 | 48.3 | #178 | |
| 3 | Llama 4 Maverick | 21.4 | #1110 | |
| 4 | Qwen 3 32B | 11.8 | #972 | |
| 5 | Gemma 3 27B | 11 | #1663 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=postcondbench-correctness · How It Works · Data refreshed daily, snapshot 2026-11-21.