Apple researchers propose SELFCOMPACT, letting LLM agents decide when to compact context

Apple 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.