Apple proposes SELFCOMPACT, letting LLM agents decide when to compact context
Overview
Apple Machine Learning Research and Johns Hopkins University researchers propose SELFCOMPACT, a scaffold that lets a model decide when and how to compact its agent context, using a compaction tool and a lightweight rubric.
The authors say it matches or exceeds fixed-interval summarization at a lower token cost, and that it needs no fine-tuning or external supervision.
The authors report gains over a no-summarization baseline of up to 18.1 points on math and 5 to 9 points on agentic search, with per-question cost 30 to 70% lower. These figures come from the authors' own report.
Written by AI from the articles below · updated Oct 12, 7:53 PM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
- Apple Machine Learning ResearchOfficialApple researchers propose SELFCOMPACT, letting LLM agents decide when to compact context
AIApple Machine Learning Research and Johns Hopkins University researchers propose SELFCOMPACT, a scaffold that lets models decide when to compact their context using a compaction tool plus a lightweight rubric. The authors report it matches or exceeds fixed-interval summarization at a fraction of the token cost, improving over a no-summarization baseline by up to 18.1 points on math and 5–9 points on agentic search at 30–70% lower per-question cost. The method requires no fine-tuning or external supervision.
Heat trend
Not enough continuous observations to show a trend yet.