Make context compaction a trained policy, not an inference-time heuristic — CompactionRL lifts SWE-bench up to +7 points

arXiv 2607.05378·high signal

CompactionRL (Tsinghua, arXiv 2607.05378, July 6 2026) folds summarization into RL rollout collection so the agent learns WHAT to keep when it compresses, optimizing the summary and the task under one final reward. Under fixed 64k–80k windows it adds +5.5 to +7.0 points on SWE-bench Verified and +3 to +6.8 on Terminal-Bench versus heuristic compaction. For builders: if you run long-horizon coding agents, the summary step is a tunable component — treat it as policy (token-level PPO loss, cross-trajectory credit assignment), not a fixed 'summarize the last N turns' prompt.