外观
论文HF Daily Papers · arXiv 2610.349212 天前
PhysEvolve框架：仿真验证驱动VLM物理推理自进化
本文由 AI 根据公开内容整理，可能有误，请以原文为准：PhysEvolve: Verifier-grounded Self-evolution for Physical Reasoning with Executable Worlds（HF Daily Papers · arXiv 2610.34921）
核心内容
针对视觉语言模型（VLM）在具身物理因果推理中因缺乏反事实训练数据而表现不佳的问题，研究团队提出PhysEvolve验证器驱动自进化框架。该方法以Qwen3-VL-8B-Instruct为基座，无需人工干预数据，通过在PhysBench、CausalPhys和ContPhy三大基准测试中的自主迭代训练，实现了物理推理能力的显著增强，其中PhysBench投掷任务提升8.8分，ContPhy反事实移除与目标陷阱任务分别提升7.4分和24.4分。
技术上，PhysEvolve构建了“模型提议-仿真验证”闭环：共享VLM生成可执行世界程序与证据模式，确定性物理引擎冻结有效环境并模拟基线与反事实分支，合成离散视觉任务。系统利用精确验证器反馈，通过GRPO算法更新策略并自适应调整课程难度。实验表明，该机制使模型掌握了交互敏感型因果原则，并能零样本迁移至布料、流体及软体等未见过的复杂真实世界动力学场景中。
关键看点
- 自进化闭环：VLM提议可执行世界程序，物理引擎模拟反事实分支并合成任务，形成无监督训练数据流
- 算法优化机制：采用GRPO算法结合精确验证器反馈更新共享策略，同步构建基于物理意图的自适应课程
- 量化性能增益：Qwen3-VL-8B在PhysBench投掷任务涨8.8分，ContPhy目标陷阱任务大幅提升24.4分
- 泛化迁移能力：因果推理原则实现零样本迁移，有效覆盖布料、流体、软体等复杂动力学及真实视频场景
工程与产业观察
该工作验证了“仿真器作为验证真理源”在具身大模型训练中的有效性，解决了纯视频数据缺乏反事实因果的痛点。对于机器人算法工程师而言，利用确定性物理引擎自动生成带标签的反事实数据并结合RLHF类算法（GRPO），是低成本提升VLM物理常识与操作规划能力的可行技术路径，对Sim-to-Real数据工程具有重要参考价值。
英文摘要 (Abstract)
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 , 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, 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.
Hugging Face Daily Papers 热度：0 upvotes 作者：Jinyuan Hu, Yilei Jiang, Chen Tang, Minghong Cai, Xinran Liu, Xiangyu Yue
来源
本页事实对应的公开页面，以原文为准。