Sample-Efficient Post-Training for LEGO Spatial-Physics Reasoning

In three linesLLM-based LEGO assembly generation suffers from PhysHack: physically valid but geometrically misaligned structures. Authors propose PVPO, a sample-efficient RL method coupling physical feasibility with voxel-space geometric rewards. Results: improved semantic alignment, structural stability, and calibration across model backbones.

Summary generated by Claude — human-verified