PhysEvolve trained a VLM on self-generated executable physics worlds with no human intervention data, improving ContPhy Goal-Pit accuracy by 24.44 points to 44.44% and surpassing even Qwen3-VL-32B's 26.7%
AspireAI community · arXiv (Shanghai Jiao Tong / HKUST)
PhysEvolve, an arXiv paper submitted October 26, 2026, tackles physical reasoning in vision-language models, where observational video captures only the factual trajectory that occurred and grounded counterfactual training data is scarce. The framework lets a shared VLM propose executable world programs and evidence schemas, freezes each valid world in a deterministic physics engine, simulates matched baseline and body-elided counterfactual branches, and synthesizes discrete answer-bound visual tasks. Exact verifier feedback from task solving then updates the shared policy via GRPO alongside an adaptive curriculum over physical intents. Starting from Qwen3-VL-8B-Instruct, the authors report PhysBench Throwing improving by 8.8 points, ContPhy Counterfactual Removal by 7.38 points (51.64% versus 44.26%) and ContPhy Goal-Pit by 24.44 points (44.44% versus 20.00%), with the Goal-Pit result surpassing even Qwen3-VL-32B's 26.7%. Gains concentrate on interaction-sensitive and altered-world questions, and the authors demonstrate zero-shot transfer to continuous media including cloth, fluid, rope and soft bodies. An ablation makes the case for the verifier explicitly: without Proposer-role updates the policy suffers severe representational collapse, with accuracy dropping to 15.01% on CausalPhys and 17.85% on ContPhy. The limitation is that the counterfactuals come from a deterministic simulator, so fidelity is bounded by how closely that simulator matches real physical behaviour.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.