CLEVER
收藏资源简介:
CLEVER是一个高质量的精选基准,包含161个编程任务,用于在Lean中生成端到端验证代码。每个问题都包括生成与保留的真实规范匹配的规范的任务,以及生成证明满足此规范的Lean实现的任务。与以前的基准不同,CLEVER避免了测试用例监督、LLM生成的注释以及可能泄露实现逻辑或允许空解决方案的规范。所有输出都使用Lean的类型检查器进行后验验证,以确保机器可检查的正确性。CLEVER用于评估基于最先进语言模型的几个少样本和代理方法。这些方法都难以实现完全验证,从而确立了它在程序合成和形式推理方面的挑战性前沿基准。我们的基准可以在GitHub和HuggingFace上找到。所有我们的评估代码也在线可用。
CLEVER is a high-quality curated benchmark comprising 161 programming tasks for generating end-to-end verified code in the Lean theorem prover. Each problem includes two core tasks: generating a specification that matches the preserved ground-truth specification, and generating a Lean implementation that is formally proven to satisfy this specification. Unlike prior benchmarks, CLEVER avoids test case supervision, LLM-generated annotations, and specifications that could leak implementation logic or allow for trivial or vacuous solutions. All outputs are verified post-hoc using Lean's type checker to ensure machine-checkable correctness. CLEVER is employed to evaluate several few-shot and AI Agent methods based on state-of-the-art language models. All of these methods struggle to achieve full verification, establishing CLEVER as a challenging cutting-edge benchmark for program synthesis and formal reasoning. Our benchmark is available on both GitHub and HuggingFace, and all of our evaluation code is also publicly accessible online.

- 1CLEVER: A Curated Benchmark for Formally Verified Code Generation德克萨斯大学奥斯汀分校, 加州理工学院 · 2025年



