Paper page - PhysEvolve: Verifier-grounded Self-evolution for Physical Reasoning with Executable Worlds

Source: https://huggingface.co/papers/2610.34921

Abstract

Physevolve trains vision-language models in causal physical reasoning by using a physics engine to validate self-generated executable worlds and counterfactual simulations, improving performance on dynamics benchmarks without human data.

Physical reasoningunderpins how embodied systems understand and act in the physical world, yetvision-language models(VLMs) remain unreliable on questions about dynamics and physical causality. Progress is critically constrained by the scarcity of groundedcounterfactualtraining data, as observational videos capture only the factual trajectory that occurred. Theself-evolvingparadigm 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-groundedself-evolvingframework that trains VLMs forcausal physical reasoningfrom model-proposed executable worlds and simulation-authored tasks. At each round, a shared VLM proposesexecutable world programsand evidence schemas; a deterministicphysics enginefreezes each valid world, simulates matched baseline and body-elidedcounterfactualbranches, and synthesizes discrete, answer-bound visual tasks. Exact verifier feedback from task solving updates both the shared policy viaGRPOand an adaptive curriculum over physical intents. Starting from Qwen3-VL-8B-Instruct, achieves targeted gains acrossPhysBench, CausalPhys, andContPhywithout human intervention data, improvingPhysBenchThrowing by 8.8 points andContPhyCounterfactualRemoval 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.

View arXiv pageView PDFProject pageGitHub2Add to collection

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2610.34921 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2610.34921 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2610.34921 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.