Skip to content- Sources: primary
- Summary: Context management in long-horizon coding agents is usually a harness heuristic outside the model, and AutoCompact folds the decision into the policy: when to compact, what working state to keep, and how to resume. Training data comes from running the base agent, having a judge review its compaction decisions, summaries, and post-compaction actions, then replacing flawed outputs with corrected ones before execution, with supervised fine-tuning followed by joint RL on coding and compaction under task-success rewards. The paper was posted 2026-10-01 and reports absolute pass-rate gains of 9.2% on SWE-bench Verified and 5.0% on SWE-PolyBench Verified, with no independent reproduction.
- Why it matters: The reported gains hold across inference budgets, which suggests context compaction is better trained into the model than configured in the harness.