LiveFMBench - Pre2025: leaderboard

The historical split of LiveFMBench (2026): 270 C programs from earlier ACSL benchmarks and GitHub, for which a model must write ACSL specifications (preconditions, postconditions, loop invariants) that let Frama-C prove the embedded assertions. The score is pass@1 (%) over 32 samples under direct prompting, counting a sample only when the proof succeeds and the model left the original code and assertions unchanged; higher is better.

Metric: Pass@1 (%) after the faithfulness check on LiveFMBench-Pre2025 (270 C programs from AutoSpec, FMBench and GitHub), direct prompting (instructions plus the C program), 32 samples at temperature 0.7 and top-p 0.95, pass@1 estimated from the 32 samples with the unbiased estimator; a sample counts only when Frama-C/WP (Alt-Ergo and Z3) proves the program and its code and assertions are unchanged (faithfulness check); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 11 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Qwen 3 235B A22B 2507 (Thinking)34.92#471
2DeepSeek R1 052834.69#432
3DeepSeek V3.1 TerminusThinking30.52#442
4DeepSeek V3.2Thinking30.31#425
5Kimi K2 090527.49#578
6Qwen 3 32B27.44#983
7GLM-4.627.08#496
8Qwen 3 Coder 480B A35B Instruct27.04#655
9DeepSeek V3 (0324)19.47#730
10Qwen 3 235B A22B 2507 Instruct11.54#625
11Llama 3.3 70B Instruct4.27#1394

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=livefmbench-pre2025 · How It Works · Data refreshed daily, snapshot 2026-11-22.