CompactionRL: Reinforcement Learning with Context Compaction for Long-Horizon Agents
Long-horizon agentic LLMs are increasingly limited by finite context windows, as extended interaction trajectories can exceed the maximum context length before a task is completed. Context compaction offers a natural solution by summarizing previous interaction states and continuing the rollout under a compressed context, but incorporating compaction into reinforcement learning remains underexplored. We propose CompactionRL, a reinforcement learning strategy to train long-horizon agentic LLMs with context compaction. Our approach jointly optimizes task execution and summary generation with tok
Lineage graph
Paper → model → repo connections mined from source citations (Tier-1 exact match).
Why these links exist
- Linked via arxiv authorYujiang Li →
CompactionRL: Reinforcement Learning with Context Compaction for Long-Horizon Agents
- Linked via arxiv authorZhenyu Hou →
CompactionRL: Reinforcement Learning with Context Compaction for Long-Horizon Agents
- Linked via arxiv authorYi Jing →
CompactionRL: Reinforcement Learning with Context Compaction for Long-Horizon Agents
- Linked via arxiv authorJie Tang →
CompactionRL: Reinforcement Learning with Context Compaction for Long-Horizon Agents
- Linked via arxiv authorYuxiao Dong →
CompactionRL: Reinforcement Learning with Context Compaction for Long-Horizon Agents