PhysEvolve: Verifier-grounded Self-evolution for Physical Reasoning with Executable Worlds
作者: Jinyuan Hu, Yilei Jiang, Chen Tang, Minghong Cai, Xinran Liu, Xiangyu Yue
分类: cs.AI
发布日期: 2026-10-26
备注: Preprint
💡 一句话要点
提出PhysEvolve以解决物理推理中的数据稀缺问题
🎯 匹配领域: 支柱九:具身大模型 (Embodied Foundation Models)
关键词: 物理推理 视觉语言模型 自我进化 因果推理 动态模拟 验证者框架 复杂场景理解
📋 核心要点
- 现有的视觉语言模型在处理物理因果性和动态问题时表现不佳,缺乏有效的反事实训练数据。
- 本文提出的 extit{PhysEvolve}框架允许模型自我生成训练问题,并通过外部验证者来验证物理环境和结果。
- 实验结果显示, extit{PhysEvolve}在PhysBench、CausalPhys和ContPhy等基准上取得了显著提升,尤其是在复杂动态场景下的零-shot迁移能力。
📝 摘要(中文)
物理推理是具身系统理解和行动于物理世界的基础,但现有的视觉语言模型(VLMs)在动态和物理因果性问题上仍然不可靠。由于观察视频仅捕捉到发生的事实轨迹,缺乏基于真实情况的反事实训练数据,进展受到限制。为此,本文提出了 extit{PhysEvolve},一个基于验证者的自我进化框架,能够从模型提出的可执行世界和模拟生成的任务中训练VLMs进行因果物理推理。实验表明, extit{PhysEvolve}在多个基准上取得了显著提升,尤其是在与交互敏感和改变世界的问题相关的任务上。
🔬 方法详解
问题定义:本文旨在解决视觉语言模型在物理推理中的不可靠性,尤其是缺乏有效的反事实训练数据的问题。现有方法依赖于固定的数据集,限制了模型的学习能力。
核心思路: extit{PhysEvolve}框架通过自我进化的方式,允许模型生成可执行的世界程序和证据模式,并利用外部验证者来验证这些生成的环境及其物理结果。
技术框架:该框架包括多个模块:共享的视觉语言模型(VLM)提出可执行的世界程序,确定性物理引擎对有效世界进行冻结和模拟,并合成离散的视觉任务。每轮的反馈通过GRPO更新共享策略和物理意图的自适应课程。
关键创新: extit{PhysEvolve}的创新在于引入了验证者的概念,使得模型能够在没有人工干预的情况下自我生成和验证训练数据,这与传统方法依赖于静态数据集的方式有本质区别。
关键设计:在设计上, extit{PhysEvolve}使用了确定性物理引擎来模拟物理环境,并通过精确的验证反馈来更新模型策略,确保模型能够适应复杂的物理动态场景。
🖼️ 关键图片
📊 实验亮点
实验结果显示, extit{PhysEvolve}在PhysBench的Throwing任务上提升了8.8分,在ContPhy的Counterfactual Removal和Goal-Pit任务上分别提升了7.4分和24.4分,尤其在交互敏感和改变世界的问题上表现突出,展示了其在复杂动态场景中的零-shot迁移能力。
🎯 应用场景
extit{PhysEvolve}的研究成果在机器人、自动驾驶、虚拟现实等领域具有广泛的应用潜力。通过提升物理推理能力,能够使得智能系统更好地理解和适应复杂的物理环境,从而提高其决策和操作的准确性。
📄 摘要(原文)
Physical reasoning underpins how embodied systems understand and act in the physical world, yet vision-language models (VLMs) remain unreliable on questions about dynamics and physical causality. Progress is critically constrained by the scarcity of grounded counterfactual training data, as observational videos capture only the factual trajectory that occurred. The self-evolving paradigm offers a route beyond fixed datasets by enabling models to construct their own training problems. Extending this paradigm to physics, however, requires an external source of truth that can both validate model-proposed environments and verify alternative physical outcomes under intervention. We introduce \physevolve{}, a verifier-grounded self-evolving framework that trains VLMs for causal physical reasoning from model-proposed executable worlds and simulation-authored tasks. At each round, a shared VLM proposes executable world programs and evidence schemas; a deterministic physics engine freezes each valid world, simulates matched baseline and body-elided counterfactual branches, and synthesizes discrete, answer-bound visual tasks. Exact verifier feedback from task solving updates both the shared policy via GRPO and an adaptive curriculum over physical intents. Starting from Qwen3-VL-8B-Instruct, \physevolve{} achieves targeted gains across PhysBench, CausalPhys, and ContPhy without human intervention data, improving PhysBench Throwing by $8.8$ points and ContPhy Counterfactual Removal and Goal-Pit by $7.4$ and $24.4$ points. These gains concentrate on interaction-sensitive and altered-world questions, demonstrating zero-shot transfer of causal principles to unseen complex dynamics (cloth, fluid, soft bodies) and real-world video.