AI News AI资讯 7d ago Updated 6d ago 更新于 6天前 51

Mistral AI Releases Leanstral 1.5: An Apache-2.0 Lean 4 Code Agent Model Solving 587 of 672 PutnamBench Problems Mistral AI发布Leanstral 1.5:一个Apache-2.0授权的Lean 4代码智能体模型,解决PutnamBench中672道题中的587道

Mistral AI released Leanstral 1.5, an open-weight (Apache 2.0) code agent model based on the Mistral Small 4 family, specifically designed for Lean 4 automated theorem proving. The model utilizes a Mixture-of-Experts (MoE) architecture with 119B total parameters and 6.5B active parameters per token, supporting a 256k context window and multimodal input. It achieves state-of-the-art performance, saturating miniF2F (100%) and significantly outperforming larger open-source models and proprietary co Mistral AI发布Leanstral 1.5,专为Lean 4证明助手设计的代码智能体模型,采用Apache 2.0开源协议。 模型基于Mistral Small 4家族,拥有119B参数(MoE架构,每token激活6.5B),支持256k上下文及多模态输入。 在miniF2F基准测试中达到100%饱和,PutnamBench解决587/672问题,并在FATE-H/X等代数基准上刷新SOTA。 训练包含中期训练、监督微调及CISPO强化学习,通过多轮交互和文件系统工具使用优化代理行为。 具备极强的测试时扩展能力,成本仅为同类模型的几分之一,并能验证真实Rust代码中的逻辑漏洞。

75
Hot 热度
70
Quality 质量
72
Impact 影响力

Analysis 深度分析

TL;DR

  • Mistral AI released Leanstral 1.5, an open-weight (Apache 2.0) code agent model based on the Mistral Small 4 family, specifically designed for Lean 4 automated theorem proving.
  • The model utilizes a Mixture-of-Experts (MoE) architecture with 119B total parameters and 6.5B active parameters per token, supporting a 256k context window and multimodal input.
  • It achieves state-of-the-art performance, saturating miniF2F (100%) and significantly outperforming larger open-source models and proprietary competitors like Opus 4.6 on FLTEval benchmarks.
  • Training involves a three-stage process including Reinforcement Learning with CISPO, utilizing multiturn and code agent environments to refine proofs through compiler feedback and filesystem interaction.
  • Beyond pure mathematics, the model demonstrates practical utility in software engineering by verifying code complexity (e.g., AVL trees) and detecting genuine bugs in open-source Rust repositories.

Why It Matters

This release marks a significant milestone in the intersection of formal verification and large language models, demonstrating that efficient, open-weight models can rival or exceed expensive proprietary systems in complex reasoning tasks. For AI practitioners and researchers, it provides a highly accessible, cost-effective tool for automated theorem proving and formal method applications, lowering the barrier to entry for integrating LLMs into critical software verification pipelines.

Technical Details

  • Architecture: Mixture-of-Experts (MoE) with 128 experts, activating 4 per token. Total size is 119B parameters, with 6.5B activated. It supports 256k token context length and accepts multimodal input (text and images).
  • Training Methodology: Utilizes a three-stage pipeline: mid-training, supervised fine-tuning, and reinforcement learning with CISPO. Two specific RL environments were employed: a multiturn environment for iterative proof refinement based on compiler feedback, and a code agent environment interacting with a raw filesystem and Lean language server.
  • Verification & Scaling: Correctness is verified using Mistral’s fork of SafeVerify. The model exhibits strong test-time scaling, where increasing the token budget per attempt drastically improves success rates on PutnamBench (from 44 solved at 50k tokens to 587 at 4M tokens).
  • Performance Metrics: Achieved 100% on miniF2F, 587/672 on PutnamBench, and new SOTA on FATE-H (87%) and FATE-X (34%). On FLTEval, it achieved a pass@1 of 28.9 and pass@8 of 43.2, surpassing Opus 4.6 at approximately one-seventh the cost.

Industry Insight

  • Cost-Efficiency in Formal Verification: Leanstral 1.5 proves that high-performance formal reasoning does not require exorbitant computational budgets. Its ability to outperform larger models at a fraction of the cost makes it viable for widespread adoption in industries requiring rigorous code verification, such as finance and aerospace.
  • Integration of Agentic Workflows: The success of the code agent environment highlights the importance of tools that allow LLMs to interact directly with compilers, file systems, and language servers. Developers should prioritize building agentic workflows that leverage real-time feedback loops rather than relying solely on static generation.
  • Open-Source Advantage in Niche Domains: The release underscores the competitive strength of open-weight models in specialized domains like mathematical proof and formal methods. Organizations can now deploy these models locally or via affordable APIs without licensing constraints, fostering innovation in secure and verified software development.

TL;DR

  • Mistral AI发布Leanstral 1.5,专为Lean 4证明助手设计的代码智能体模型,采用Apache 2.0开源协议。
  • 模型基于Mistral Small 4家族,拥有119B参数(MoE架构,每token激活6.5B),支持256k上下文及多模态输入。
  • 在miniF2F基准测试中达到100%饱和,PutnamBench解决587/672问题,并在FATE-H/X等代数基准上刷新SOTA。
  • 训练包含中期训练、监督微调及CISPO强化学习,通过多轮交互和文件系统工具使用优化代理行为。
  • 具备极强的测试时扩展能力,成本仅为同类模型的几分之一,并能验证真实Rust代码中的逻辑漏洞。

为什么值得看

本文展示了AI在形式化验证和复杂数学证明领域的最新突破,证明了轻量级MoE模型在特定垂直领域可达到甚至超越更大规模闭源模型的性能。对于从事自动定理证明、形式化方法或高可靠性代码开发的从业者而言,Leanstral 1.5提供了极具性价比的开源解决方案,并揭示了“代码即证明”在工程实践中的巨大潜力。

技术解析

  • 架构与规格:模型属于Mistral Small 4系列,采用混合专家(MoE)架构,总计119B参数,但每个token仅激活6.5B参数,由128个专家中的4个共同处理。支持256k长上下文窗口,输入端支持文本和图像多模态,输出为纯文本。
  • 训练流程:采用三阶段训练策略:中期训练、监督微调(SFT)以及基于CISPO的强化学习。强化学习环境包括“多轮环境”(根据编译器反馈迭代证明)和“代码代理环境”(直接在文件系统中编辑文件、运行bash命令并与语言服务器交互)。
  • 性能基准:在miniF2F验证集和测试集上均达到100%准确率;在PutnamBench上解决587道题(约$4/题,远低于竞争对手的$300+);在FATE-H和FATE-X代数基准上分别达到87%和34%,创下新纪录;在FLTEval上Pass@8达到43.2,超越Opus 4.6且成本更低。
  • 测试时扩展性:模型表现出显著的测试时缩放效应,随着每尝试的token预算增加(从50k到4M),PutnamBench的解决数量从44个线性增长至587个,验证了其在长推理路径上的有效性。
  • 代码验证案例:成功证明了AVL树实现的O(log n)时间复杂度(使用结构归纳法和TimeM单子追踪);通过Aeneas将Rust代码翻译为Lean,在57个仓库中发现11个真实Bug,其中5个未公开报告,包括一个导致溢出崩溃的关键错误。

行业启示

  • 形式化验证的工程化落地:AI不再仅限于生成代码,而是能够深入理解并验证代码的正确性属性。开发者可利用此类模型自动生成不变量、补全部分证明或进行压力测试,显著提升软件供应链的安全性。
  • 成本效益驱动的模型选择:Leanstral 1.5以较小的激活参数量和极低的单次推理成本,击败了参数量大得多的闭源模型。这表明在特定任务(如定理证明)中,经过精心调优的垂直领域小模型比通用大模型更具经济优势和实用性。
  • 代理式工作流的标准化:模型通过与文件系统、编译器后端和语言服务器的深度集成来执行任务,确立了“代码代理”的新范式。未来AI助手将更倾向于作为自主实体在开发环境中操作,而非仅仅作为聊天机器人提供建议。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Open Source 开源 Agent Agent Code Generation 代码生成 Programming 编程 Benchmark 基准测试