SELFCOMPACT: model-invoked adaptive compaction for language-model agents
SELFCOMPACT pairs a model-invoked compaction tool with a lightweight inference-time rubric that signals when to summarize or suppress compaction, letting the model decide how and when to compact accumulated agent traces. Across six benchmarks and seven models, this approach matches or exceeds fixed-interval summarization while cutting per-question token cost by 30–70%, improving math performance by up to 18.1 points and agentic-search results by 5–9 points without fine-tuning.