Compress to Remember: Learning Compact Memory via On-Policy Distillation for Long Video Generation
Published in arXiv preprint, 2026
Prediction-Aligned Context Compaction (PACC) learns to summarize past video frames into reusable memory tokens. On-policy distillation uses the same frozen generator as a teacher with full history and a student with compressed memory, updating only the compressor. PACC improves memory-event coverage and consistency on MBench while maintaining competitive quality in minute-long generation.
Recommended citation: Wu, X.*, Guo, W.*, Wang, Y.*, Feng, X., Kavraki, L. E., and Wu, Z. S. (2026). "Compress to Remember: Learning Compact Memory via On-Policy Distillation for Long Video Generation." arXiv:2609.36364.
Download Paper