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

PhysEvolve. Verifier-grounded Self-evolution for Physical Reasoning with Executable Worlds. By Jinyuan Hu and five others, from the Chinese University of Hong Kong. Published on the twenty-sixth of October.

Physical reasoning underpins how embodied systems understand and act in the physical world, yet vision-language models remain unreliable on questions about dynamics and physical causality. Progress is critically constrained by a scarcity of grounded counterfactual training data. Observational video only captures the trajectory that actually occurred, never the one that would have followed had something been different. The self-evolving paradigm, in which models construct their own training problems, offers a route beyond fixed datasets. Carrying it into physics, however, needs an external source of truth: something that can validate the environments a model proposes, and verify alternative physical outcomes under intervention.

PhysEvolve makes a physics engine that source of truth. It is a verifier-grounded self-evolving framework that trains vision-language models for causal physical reasoning, from executable worlds the model proposes and from tasks the simulation authors. The division of labour is the point. The model is free to invent worlds, but what actually happens inside them is decided by the simulator, so every answer it trains on is exact.

Each round begins with a shared VLM proposing executable world programs, together with evidence schemas. A deterministic physics engine freezes each valid world, then simulates matched branches: a baseline, and a counterfactual in which a body has been elided. From those branches it synthesizes discrete visual tasks whose answers are bound to the simulation. Exact verifier feedback from solving the tasks then updates two things at once: the shared policy, through GRPO, and an adaptive curriculum over physical intents. The same model that proposes worlds is the one that learns from them.

The matched branches are what make the data counterfactual. A baseline and its body-elided twin start from the same frozen world, so whatever differs between them is a consequence of the intervention, and a task built from the pair can ask about exactly that consequence. Because the engine is deterministic, the answer to that task is not a guess or a label someone wrote down. It is the outcome the simulation produced.

Training starts from Qwen3-VL-8B-Instruct and uses no human intervention data. PhysEvolve achieves targeted gains across PhysBench, CausalPhys and ContPhy. PhysBench Throwing improves by eight point eight points. On ContPhy, Counterfactual Removal improves by seven point four points, and Goal-Pit by twenty-four point four points. These gains concentrate on interaction-sensitive and altered-world questions, exactly the kind of question that observation alone cannot supply answers for.

Those concentrated gains carry the broader claim: zero-shot transfer of causal principles to unseen complex dynamics, namely cloth, fluid and soft bodies, and to real-world video. The counterfactual data that no camera could record is written here by a simulator, inside worlds the learner itself proposed.