Language models can't spark scientific revolutions, but world models might 语言模型无法引发科学革命，但世界模型可以

TL;DR

• Language models lack "manipulative abduction," the creative leap required to invent entirely new axioms or frameworks, which is essential for scientific revolution.
• While AI excels at deduction (logical derivation) and induction (pattern recognition), it struggles with the embodied, sensory-based intuition that drives foundational discoveries like Einstein's relativity.
• Current systems optimize within existing paradigms but cannot overthrow established theories without a detectable error signal, unlike human scientists who make leaps based on physical simulation rather than data discrepancies.
• The paper proposes action-controllable world models as a potential path forward, enabling agents to run counterfactual experiments in synthetic labs to generate new axioms through embodied interaction rather than passive pattern matching.

Why It Matters

This analysis challenges the assumption that scaling language models alone will lead to autonomous scientific discovery, highlighting a fundamental cognitive gap between statistical prediction and genuine innovation. For AI researchers, it underscores the need to move beyond pure symbol manipulation toward architectures grounded in physical simulation and embodied reasoning to achieve true breakthrough capability. The distinction also clarifies why current AI assistants excel at optimizing known processes but struggle to propose radically new theoretical frameworks.

Technical Details

• Zahavy employs Peirce's tripartite reasoning framework: deduction (rule-to-conclusion), induction (case-to-rule), and abduction (surprise-to-explanation), arguing only the latter two are accessible to LLMs while manipulative abduction remains out of reach.
• The critique centers on optimization-driven learning requiring detectable error signals; since Einstein's theory emerged during a period where Newtonian physics was overwhelmingly validated (with Mercury's anomaly attributed to a hypothetical planet), an AI would have no incentive to discard the existing paradigm.
• Embodied simulation—exemplified by Einstein's elevator thought experiment and Archimedes' bathtub insight—is identified as the source of manipulative abduction, contrasting sharply with LLMs' reliance on linguistic patterns without sensory grounding.
• Action-controllable world models (e.g., Genie) are distinguished from generative video models (e.g., Veo) by their capacity for active intervention and counterfactual experimentation, potentially creating the synthetic feedback loops needed for axiom invention.

Industry Insight

AI development should prioritize integrating physical simulation and embodied reasoning capabilities into next-generation systems rather than solely pursuing scale in language modeling, as this addresses the core bottleneck for revolutionary scientific contribution. Research funding and architectural design must shift toward creating "synthetic laboratories" where agents can manipulate environments and test hypotheses through direct interaction, mimicking the cognitive process behind foundational discoveries. This reframes the goal of AI-assisted science from automating incremental progress to enabling machines capable of conceptual paradigm shifts—a challenge requiring fundamentally different engineering approaches than current LLM training paradigms.

TL;DR

• 谷歌DeepMind的Tom Zahavy在立场论文《LLMs can't jump》中提出，语言模型缺乏创造真正新事物所需的认知机制。
• Zahavy基于爱因斯坦的理论框架，指出发现过程包括感官体验、直觉“跳跃”和逻辑推理三个步骤，而AI仅能处理其中两个环节。
• 他区分了两种形式的溯因推理（abduction）：普通溯因（从已知候选中选择最合理解释）和操纵性溯因（发明尚无语言模板的新解释），认为后者是科学发明的瓶颈。
• 当前AI系统如AlphaProof、Gemini和GPT-5已在数学推导和模式识别上达到高水平，但无法像爱因斯坦那样通过感官模拟提出全新公理。
• Zahavy建议发展具身世界模型（action-controllable world models），让AI能通过反事实实验进行主动干预，从而可能实现真正的科学创新。

为什么值得看

这篇文章对AI从业者具有重要启示意义，它指出了当前大语言模型在科学发现中的根本局限——缺乏从感官经验到抽象公理的创造性跃迁能力。对于从事基础科学研究或AI+科学交叉领域的研究者而言，理解这一瓶颈有助于更合理地评估AI在科学革命中的角色，并探索更具潜力的技术路径，如具身智能和世界模型的发展。

技术解析

• Zahavy引用哲学家Charles Sanders Peirce的推理分类法，将推理分为演绎（deduction）、归纳（induction）和溯因（abduction）。当前AI擅长前两者：演绎用于形式化证明（如国际数学奥林匹克金牌水平），归纳用于数据模式识别（如统计学习）。
• “操纵性溯因”被定义为发明一个此前不存在于任何语言模板中的因果解释，这是科学理论突破的核心。Zahavy认为AI无法完成此类操作，因为其优化依赖可检测的错误信号，而爱因斯坦提出相对论时并无数据危机。
• 以牛顿力学为例，当时仅有水星轨道微小异常，传统方法会假设隐藏行星“火神星”来拟合数据；而爱因斯坦则通过思想实验（自由落体观察者无重力感）直接重构时空概念，这种突破源于具身模拟而非误差最小化。
• 对比现有AI科学工具：Sakana的AI Scientist仅能重组已有概念，DeepMind的AlphaEvolve虽能高效优化但需明确目标函数；二者均无法在无先验框架下生成新范式。
• 潜在解决方案是构建“行动可控的世界模型”（如Genie），允许代理在仿真环境中执行反事实操作（例如切断电梯缆绳），从而形成类似人类心智的实验反馈循环，为操纵性溯因提供可能的基础。

行业启示

• AI在辅助科研流程自动化方面已具强大能力，但在原创性理论构建上仍存在本质鸿沟。未来投资应聚焦于融合感知、动作与推理的具身智能架构，而非单纯扩展语言模型的规模或上下文长度。
• 科学发现的本质不仅是数据处理，更是基于身体经验的想象力跃迁。因此，开发具备物理交互能力的虚拟实验室（synthetic lab）可能是通往下一代AI科学家的重要方向，这类系统应支持主动干预与反事实推演。
• 学术界和产业界需重新审视“AI驱动科学”的定义边界：短期来看，AI更适合加速验证阶段的工作；长期突破则需等待能够模拟真实世界因果结构的通用认知模型出现，这可能要求跨学科整合神经科学、物理学与控制论的研究成果。