PhysEvolve: CUHK trains VLM physical reasoning with physics-engine-verified self-generated worlds
Jinyuan Hu · hf · 2026-10-28
MMLab at CUHK introduces PhysEvolve, a verifier-grounded self-evolution framework training VLMs for causal physical reasoning. The pain point: VLMs are unreliable on dynamics and physical causality, and progress is constrained by scarce counterfactual training data—observational video captures only the trajectory that actually occurred.
Each round, a shared VLM proposes executable world programs and evidence schemas; a deterministic physics engine freezes valid worlds and simulates counterfactual trajectories as ground truth, verifying model-proposed environments and validating alternative outcomes under intervention. Models thus author their own training problems without human data, improving performance on dynamics benchmarks.
More from Research
- Google's AI Mode feeds uncited Knowledge Graph biographies into answers, analysis shows — dejanseo · 2026-10-28
- DFR Framework Tackles Reward Fine-Tuning in Diffusion Models Without Losing Diversity — mmbronstein · 2026-10-28
- alphaXiv ships Mobile 2.0 with AI paper reader and one-click visual blogs — datawithsuman · 2026-10-28
- Researchers screened 175 Spanish books ahead of time and found multiple AI-generated positives — TuhinChakr · 2026-10-28
- Xiaohongshu and Fudan Propose DR-Rubric: Agent-Written Rubrics as RL Rewards — jiqizhixin · 2026-10-28
- Ben Goertzel: AI Peer Review Already Beats Humans, and Human Bias Is Harder to Fix — bengoertzel · 2026-10-28