PhysEvolve trains VLMs on executable worlds for physical reasoning
Original title: PhysEvolve: Verifier-grounded Self-evolution for Physical Reasoning with Executable Worlds
From Qwen3-VL-8B-Instruct, the paper reports PhysBench Throwing up 8.8 and ContPhy Goal-Pit up 24.4 points.
PhysEvolve is a verifier-grounded self-evolving framework that trains vision-language models for causal physical reasoning using model-proposed executable worlds and simulation-authored tasks. A shared VLM proposes world programs; a deterministic physics engine simulates baseline and body-elided counterfactual branches and builds answer-bound visual tasks. Exact verifier feedback updates the policy with GRPO and an adaptive curriculum. Starting from Qwen3-VL-8B-Instruct and without human intervention data, the authors report gains on PhysBench, CausalPhys, and ContPhy, concentrated on interaction-sensitive and altered-world questions.
Unconfirmed: Reported score gains are author-measured and not independently confirmed.
- PhysBench Throwing improved by 8.8 points, the paper says.
- ContPhy Counterfactual Removal rose 7.4 points and Goal-Pit 24.4.
- Training uses GRPO and a physics engine, with no human intervention data.
- Authors claim zero-shot transfer to cloth, fluid, soft bodies, and real video.
From the original article
1 min read at huggingface.co
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.
By Jinyuan Hu, Yilei Jiang, Chen Tang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.