VCoT-Bench-Type - Lemma: leaderboard
Metric: Weighted completion accuracy (%) on the tasks with every lemma function removed of VCoT-Bench-Type (439 tasks in the sub-benchmark), Verus proof-completion tasks built by removing verification steps from solver-lifted Rust proofs (VCoT-Lift), zero-shot greedy decoding (temperature 0 except GPT-5 mini); a completion scores 1 if Verus accepts its syntax and a GPT-5 mini judge rates it semantically equivalent to the lifted ground-truth proof, 0.5 if only semantically correct, 0.25 if only syntactically correct, 0 otherwise; higher is better. Source: arxiv.org. Saturation forecast: Around October 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 | 55.48 | #105 |
| 2 | Claude Sonnet 4.5 | 51.54 | #138 |
| 3 | DeepSeek V3.2 | 44.35 | #198 |
| 4 | Gemini 3.1 Pro (Preview) | 42.81 | #54 |
| 5 | Claude Haiku 4.5 (20251001) | 42.64 | #251 |
| 6 | Gemini 3 Flash (Preview) | 41.61 | #78 |
| 7 | DeepSeek R1 0528 | 38.53 | #217 |
| 8 | GPT-5 Mini (2025-08-07) | 36.3 | #165 |
| 9 | GPT-OSS-20B | 26.2 | #499 |
| 10 | Qwen 3 8B (Non-reasoning) | 23.63 | #667 (Qwen 3 8B) |
| 11 | Qwen 3 8B (Thinking) | 22.43 | #667 (Qwen 3 8B) |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=vcot-bench-type-lemma · How It Works · Data refreshed daily, snapshot 2026-10-11.