AI Skills AI技能 6h ago Updated 3h ago 更新于 3小时前 46

The Physicist and the Frustrated Machine 物理学家与受挫的机器

Nobel laureate Giorgio Parisi and Francesco Zamponi utilized Claude (Sonnet 4.6 and Opus 4.7) to prove the identity $a + b = 1$ for critical exponents in jamming transition, a result previously observed numerically but unproven. The collaboration highlights a shift in LLM utility from simple retrieval to complex compositional reasoning, where the model synthesizes existing theoretical frameworks rather than retrieving pre-existing proofs. Hallucinations are reframed not as malfunctions but as st 诺贝尔物理学奖得主Giorgio Parisi与LLM协作,利用Claude模型证明了jamming过渡中的临界指数恒等式a+b=1。 该证明未存在于训练数据中,而是通过模型对已有物理文献的重新组合与推导得出,体现了LLM的泛化而非记忆能力。 人类专家通过提供高精度数值验证和C++代码作为约束条件,构建了“框架”,从而引导模型生成正确的解析证明。 文章提出幻觉并非单纯的故障,而是当上下文约束不足时,模型基于统计可能性进行的插值填充。 区分了两种错误模式:“缺失片段”(可通过增加信息解决)和“冲突片段”(源于训练数据矛盾,需人工干预)。

65
Hot 热度
75
Quality 质量
60
Impact 影响力

Analysis 深度分析

TL;DR

  • Nobel laureate Giorgio Parisi and Francesco Zamponi utilized Claude (Sonnet 4.6 and Opus 4.7) to prove the identity $a + b = 1$ for critical exponents in jamming transition, a result previously observed numerically but unproven.
  • The collaboration highlights a shift in LLM utility from simple retrieval to complex compositional reasoning, where the model synthesizes existing theoretical frameworks rather than retrieving pre-existing proofs.
  • Hallucinations are reframed not as malfunctions but as statistical interpolation occurring when contextual constraints are insufficient, contrasting with "conflicting pieces" arising from contradictory training data.
  • The success relied on a specific prompting strategy: establishing rigorous numerical constraints and verification frames via C++ code generation before asking for the analytic proof, thereby limiting the model's search space.
  • Human oversight remains critical for verifying consistency, pruning obscure logic, and resolving internal contradictions within the model's generated superpositions.

Why It Matters

This case study provides empirical evidence that Large Language Models can perform genuine mathematical discovery and synthesis when guided by precise, multi-stage constraints, moving beyond pattern matching into logical derivation. It offers a practical framework for researchers to mitigate hallucinations by structuring interactions that prioritize constraint satisfaction over open-ended generation, effectively turning the LLM into a collaborative reasoning engine rather than just a knowledge base.

Technical Details

  • Model Architecture: The proof was generated using Anthropic’s Claude models, specifically Sonnet 4.6 and Opus 4.7, indicating that current state-of-the-art models possess sufficient capacity for high-level symbolic manipulation and theorem proving.
  • Prompting Strategy: The interaction consisted of forty prompts in a single session. The initial phase focused on numerical verification, requiring the model to write C++ code to solve differential equations and confirm the conjecture to high precision. This established a "frame" of verified facts before the analytic proof was requested.
  • Constraint Saturation: By saturating the context with numerical truths and specific constraints (including unpublished findings by the authors), the model’s generative space was narrowed, making the correct logical path statistically dominant and reducing the likelihood of hallucinated inconsistencies.
  • Human-in-the-Loop Verification: The authors actively reviewed the output, identifying inconsistencies in early versions which the model self-corrected. They also manually pruned unnecessary or obscure sections, demonstrating a hybrid workflow where AI generates candidates and humans curate and validate.

Industry Insight

  • Shift to Constraint-Based Prompting: Practitioners should move away from direct question-answering for complex tasks. Instead, design workflows that first establish ground-truth constraints and numerical validations to "pin the frame," significantly improving the coherence of subsequent analytical outputs.
  • Redefining Hallucination Management: Understanding hallucination as a lack of constraints rather than mere error allows for better architectural solutions. Investing in tools that help users inject precise, verifiable context (like code execution or numerical solvers) into the prompt chain can drastically reduce spurious completions.
  • Collaborative Discovery Models: The future of scientific research using AI lies in iterative, multi-turn collaborations where the AI acts as a reasoning partner capable of synthesizing disparate theories, provided the human expert provides the rigorous scaffolding and final validation.

TL;DR

  • 诺贝尔物理学奖得主Giorgio Parisi与LLM协作,利用Claude模型证明了jamming过渡中的临界指数恒等式a+b=1。
  • 该证明未存在于训练数据中,而是通过模型对已有物理文献的重新组合与推导得出,体现了LLM的泛化而非记忆能力。
  • 人类专家通过提供高精度数值验证和C++代码作为约束条件,构建了“框架”,从而引导模型生成正确的解析证明。
  • 文章提出幻觉并非单纯的故障,而是当上下文约束不足时,模型基于统计可能性进行的插值填充。
  • 区分了两种错误模式:“缺失片段”(可通过增加信息解决)和“冲突片段”(源于训练数据矛盾,需人工干预)。

为什么值得看

这篇文章为AI在科学发现中的应用提供了极具深度的案例研究,展示了如何结合领域专家的直觉与LLM的计算能力解决长期未决的科学问题。它揭示了当前大模型在复杂逻辑推理中的局限性及突破路径,即通过结构化约束而非单纯的信息投喂来引导生成,对科研工作者优化人机协作流程具有重要参考价值。

技术解析

  • 协作模式:采用“人类定义约束+LLM生成证明+人类验证修正”的闭环流程。研究人员首先让模型进行数值计算并生成C++代码以验证猜想,随后才要求提供解析证明,确保上下文包含足够的确定性事实。
  • 幻觉的新定义:将幻觉定义为“在约束耗尽时的泛化操作”。当上下文未能充分限定解空间时,模型会填充统计上最合理但全局错误的片段;反之,若约束足够(如数值验证结果),模型能收敛到唯一正确解。
  • 两类失败模式
    1. 缺失片段:训练数据未覆盖相关区域,导致模型插值。此类问题可通过RAG或长上下文缓解。
    2. 冲突片段:训练数据内部存在矛盾(如不同理论来源的冲突),导致模型权重处于叠加态。此类问题无法仅靠增加信息解决,需要外部知识或人工甄别。
  • 模型版本:使用了Claude Sonnet 4.6和Opus 4.7,其中Opus 4.7在最小监督下完成了核心推导,Sonnet 4.6用于后续步骤的微调。

行业启示

  • 从“检索”转向“约束工程”:在科学和复杂推理场景中,单纯提供背景知识(RAG)不足以消除幻觉,关键在于通过中间步骤(如数值验证、代码生成)构建严密的逻辑框架,限制模型的搜索空间。
  • 人机协作的重心转移:AI的角色应从“答案提供者”转变为“假设生成器”或“组合引擎”,而人类专家的核心价值在于定义问题边界、提供关键约束以及最终的真伪判定,特别是在处理训练数据中存在内在矛盾的知识领域。
  • 重视隐性知识的显性化:模型无法访问人类未发表的直觉或验证结果(如文中提到的非正轮廓节点验证)。在利用LLM进行前沿探索时,必须将专家的隐性约束显性化地输入上下文,才能激发模型的正确组合能力。

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。

LLM 大模型 Research 科学研究 Evaluation 评测