Researchers train a coding agent to decide when to compact its context

TL;DR

• On 1 October researchers from Singapore Management University, Nanyang Technological University and Harvard posted AutoCompact, which trains a coding agent to decide when to compact its context.
• On SWE-bench Verified the trained Qwen3-Coder-30B-A3B-Instruct solved 39.6% of tasks, up from 30.4% for the base model.
• The paper names no code release.