AutoCompact: learning when to compact context

TL;DR. Coding agents usually compact (summarize and drop old history) when the context window hits a fixed threshold. AutoCompact (SMU, NTU, Harvard; arXiv 2610.02163) adds a compact() action and trains the agent to decide when to call it, what working state to keep, and how to continue afterward. Training data comes from running the base agent while a judge reviews each compaction decision, summary and post-compaction action, and replaces flawed ones before they execute, so every trajectory continues from corrected state. SFT on those trajectories, then RL with task-success reward trains coding and compaction jointly. Result: +9.2 points on SWE-bench Verified, +5.0 on SWE-PolyBench Verified, and the gain holds even with a 256K window that never overflows. Compaction helps as a thinking tool, not only as overflow control.

flowchart LR
  A["Agent step<br/><small>inspect, edit, test</small>"] --> D{"compact()?<br/><small>agent decides</small>"}
  D -->|no| A
  D -->|yes| S["Working-state summary<br/><small>findings, code, todo</small>"]
  S --> J["Judge<br/><small>fixes bad choices</small>"]
  J --> C["Continue<br/><small>from corrected state</small>"]
  C --> T["SFT then RL<br/><small>task-success reward</small>"]
  classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
  classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
  classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
  classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
  class A input
  class D,J loop
  class S core
  class C,T exit
  linkStyle 1 stroke:#f08c00,stroke-width:2px

Key findings

• Compaction replaces earlier history with a model-written # Auto Context Summary; the original task stays in context.
• The three linked decisions (timing, summary content, following the summary) are trained together. Prior approaches trained on summaries alone, so agents often ignored the summary and repeated searches.
• +9.2 / +5.0 absolute pass-rate gains over the base model, across every tested inference budget, with a 16K window (overflow forces fallback compaction) and a 256K window (never overflows).
• The authors frame it as model-harness co-design: the harness supplies the mechanism, the model learns when to invoke it.

How this relates to prior wiki pages

• Third paper in four days moving context policy into the model. Context Language Models (10-01) let the model edit its context as a file (+11.4% BrowseComp-Plus, 21.5% fewer FLOPs). FOCUS (10-03) chose which interaction units to keep, training-free, cutting peak context 48%. AutoCompact adds the learned timing. Elvis Saravia's framing of the series, "AutoHarness, then AutoContext, now AutoCompact", names the pattern: harness functions migrating into weights.
• Pairs with KV-streams (10-03), which made dropping turns cheap in the live vLLM cache during agent RL (~2x training). AutoCompact decides when to drop; KV-streams makes the drop free of re-prefill. Neither paper uses the other.
• Confirms the "state beats history" result from yesterday's Berkeley paper (a frontier model scored 9 of 40 on long agent logs, 40 of 40 when given current state): the 256K no-overflow gain says long stale history hurts even when it fits.
• Updates agent-memory and agent-harness-engineering.

Gaps

• One base model family, coding only. Robustness across harnesses (Claude Code style vs mini-SWE-agent) is untested; the multi-harness RL result of 10-03 (same weights, 62% vs 33% by harness) suggests it matters.
• No accounting of the cache cost: every compaction invalidates the prefix cache. With cache reads 90%+ of agent input tokens, frequent learned compaction could raise the bill even as it raises accuracy.

Raw source: X Following feed, 2026-10-03 (@omarsar0); paper linked above.