EasyProver用递归智能体攻克Lean验证
AI 总结
论文作者提出EasyProver,这是一种结合递归智能体、隔离上下文、异步搜索、并行策略探测和弹性详述服务的Lean证明系统。使用未经额外训练的前沿模型,作者报告其证明了672道PutnamBench题中的669道、全部6项IMO 2026形式化任务、600个NTP4VC条件中的597个,以及全部189项VERINA任务。作者还推出Aegis,生成2045对经过差异验证的Rust-Lean程序和11533条候选规格,可用于评测其他证明器及生成强化学习奖励。
深度见解
EasyProver将智能体搜索与形式化验证相结合,Aegis则为AI定理证明器提供了面向真实代码的新评测资源。
Computer Science > Computation and Language
arXiv:2610.38174 (cs)
[Submitted on 28 Oct 2026]
Title:Scaling Lean Theorem Proving for Real-World Code Verification
Authors:Debargha Ganguly, Divya Raghunathan, Dejan Jovanović, Vikash Singh, Soonho Kong, Aaron Tomb, Aman Goel, Ferhat Erata, Sam Bayless
View a PDF of the paper titled Scaling Lean Theorem Proving for Real-World Code Verification, by Debargha Ganguly and 8 other authors
#EasyProver#Aegis#Lean#Aeneas#Rust#PutnamBench#NTP4VC#VERINA