ZipRL: Adaptive Multi-Turn Context Compression with Hindsight Response Replay

ZipRL is a novel adaptive compression framework for Large Language Models (LLMs) in multi-turn agent tasks. It combines multi-granularity compression with Hindsight Response Replay (HRR) to densify training signals during Reinforcement Learning from Verifiable Rewards (RLVR) optimization. The approach aims to balance information retention and token efficiency, outperforming uniform compression methods in task-relevant utility. The framework uses coarse-to-fine prompts for macro-compression and integrates HRR into GRPO via generalized advantage. The research is published on arXiv (2605.28069v1).