AutoCompact: Learning When to Compact Context in Long-Horizon Coding Agents

简体中文

L1 diagnostic of the in-place three-field correction contract; no coding-agent training or SWE container run.

Paper information

Field	Details
Paper	arXiv v1
Institution	The paper's first page does not list an affiliation for the first author.
First public release	2026-10-01 (arXiv v1)
Author code	No public author implementation found as of 2026-10-03.
Local adapter or method key	autocompact
Local reproduction code	src/auto_research/agent_research/latest_20261003.py

Paper method and evidence

A judge assesses when to compact, what the working summary should say, and what the next action should be. The corrected output is executed directly in the environment, then learned with SFT and outcome RL.

flowchart LR
  I[Decision, summary, and next action] --> J[Judge and correct all three fields]
  J --> E[Execute corrected output]
  E --> T[SFT and outcome RL]

Original paper figure

Figure 1 shows the overall method and its components. From the original paper; copyright remains with the authors.

Core equation

Apply \((d,s,a)\leftarrow JudgeCorrect(d,s,a)\) before executing.

Reported results

The paper reports +9.2 points on SWE-bench Verified and +5.0 on SWE-PolyBench Verified.

Local reproduction

Compare with the mechanism disabled/default state. The three-seed diagnostic is diagnostic_only=true; there is no formal relative gain or capability ranking.