On-Policy Learning

Self-Distillation with On-Policy Sampling for Multi-Turn Image Editing

MT-OPSD trains models on their own edit states, and LME-Bench tests ten-turn sessions to reduce recursive fail…