Tag
微软研究院及其合作者通过人类主导的智能体研究流程，解决了存在31年之久的KLS猜想，确立了对数凹分布的普适几何原理，并经独立研究者验证与机器形式化检验确认，为高维采样与随机算法效率理解扫清重大理论障碍。
A Purdue paper introduces a bounded, online step-level repair module for Lean 4 theorem provers that feeds compiler error messages to a pretrained repair model instead of discarding failed tactics, achieving up to a 32% relative gain in solve rate on the LeanDojo benchmark and nearly doubling performance on Lean Workbook.
Maxwell is a new model that dynamically rewrites the binaries of running AI programs like Whisper and LTX-2 at runtime to speed them up, using formal verification to confirm the optimized version preserves correctness.
doubleAI 发布 Maxwell，首个“代码执行模型”，可在程序运行时重写二进制以适配特定硬件与数据，并通过形式化证明保证优化等价，带来 38× 数据管线加速与 4.25× AI 推理内核加速。
Programming Language Foundations in Agda (PLFA) 是一本使用证明助手 Agda 讲解编程语言理论的开源教材，涵盖逻辑基础、操作语义与指称语义，并列出了多门采用该书授课的高校课程。
中国研究团队借助上海AI实验室自研的Intern-S2科学大模型，完整证明了困扰国际数学界30余年的三维Ghys猜想，随后由北京大学团队完成形式化机器验证，验证代码已全部开源。
Basis researchers used LLMs to write formal semantics for a small programming language and proved in Lean that the interpreter conforms to those rules after compilation to RISC-V, treating the existing interpreter as ground truth to overcome the specification bottleneck in verified software.
介绍 SentryWard：一个开源的 GPU kernel 形式化验证工具，可证明同步安全、无数据竞争与资源生命周期正确性，并通过系统性移除冗余同步使 NVIDIA B200 上的 kernel 最快提速 1.4 倍。
Lean v4.35.0 now ships with multiple verified kernels, including con-leche (the first Lean-written kernel with a formal consistency proof, implemented by Claude) and its Rust-refined counterpart con-ron, plus a new `lake check` command for rechecking developments externally. The move follows AI-discovered kernel soundness bugs and aims to eliminate implementation-level unsoundness without performance penalties.
介绍 IneqAgent，一个将 LLM 推理与符号计算结合的智能体框架，用于 Lean 4 中的不等式形式化证明，采用 lemma 级蒙特卡洛树搜索协调引理构造与符号后端选择，在四个基准上达到 94.09% 的成功率。
该论文研究 Verus/Rust 形式验证工作流中，正确性反馈如何反而抑制规格完整性审计所需的证据，发现独立无反馈探索比带纠正反馈的循环产生多一倍的认证反例，并在 NeurIPS 2026 AI for Verifiable Coding 工作坊被接收。
QuixiAI 开源了 Sovereign Protocol Language (SPL)，一个基于 Lean 4 的形式化库，为"主权代理协议"(Protocol of Sovereign Agency) 提供规范性语义，核心包括 Duty、Power、Immunity 等原语以及 Law 3 非干扰实验，全部定理经内核验证、无公理依赖。
Materialize describes how it reworked its `persist` metadata store to build a linearizable log using PostgreSQL's READ COMMITTED isolation instead of SERIALIZABLE, tripling write throughput after a metastable latency incident, and validated correctness by modeling PostgreSQL's behavior in Lean 4 to prove linearizability.
A GitHub repository presenting machine-checked Coq proofs of new Busy Beaver champions, including a 10-state machine whose score exceeds Graham's number (f_{ω+1} level in the fast-growing hierarchy), beating the previous BB(8), BB(9) and BB(10) champions with exact scores verified without axioms.
Creusot 0.14.0 发布，这是 Rust 的形式化验证工具，新增了对偏序（partial orders）的支持以及为快照（snapshots）生成验证条件（VCs）的功能。
LeanPRM提出利用Lean证明助手自动为自然语言数学证明生成步骤级正确性标签，无需人工标注即可训练过程奖励模型（PRM）。Qwen3-4B验证器在该数据上训练后，步级标签准确率从48.2提升至74.7，并在IMO级问题的最佳-of-N选择和多个外部验证基准上超越PRM800K训练的验证器。
VEPA is a framework for verified evolution of formal guardrails for LLM agents, using LLM-proposed symbolic policy patches validated by a deterministic verifier and an evidence-aware authority layer, outperforming baselines on agent safety and security benchmarks.
LWN 报道了 Eurydice 项目——一个将 Rust 代码编译为结构清晰、可读的 C 代码的编译器工具，面向高可靠性软件验证场景，已用于后量子密码学例程的编译，由 Inria 与 Microsoft 团队维护的 Aeneas 项目的一部分。
OpenAI used its AI systems to help produce a formal proof related to the Navier-Stokes equations, but the model mistranslated parts of the mathematics into code during the formalization process.
A hands-on blog post comparing four theorem provers—Lean, Isabelle/HOL, HOL4, and Agda—by formalizing Euclid's proof of the infinitude of primes in each, with opinionated rankings on usability, interactivity, and enjoyment.