SELFCOMPACT lets models compact own context, up 18.1 points on math

Long agent traces built from chains of thought and tool calls accumulate stale content that anchors later generations and eventually outgrows the context window. Existing scaffolds handle this with fixed-interval compaction triggered at a token threshold, which pays no attention to trajectory structure and risks discarding partial results mid-derivation or mid-search.

SELFCOMPACT addresses that by pairing two inference-time elements: a compaction tool the model invokes to summarize accumulated context, and a lightweight rubric that specifies when to fire and when to suppress. The authors note both are needed—the tool alone is unevenly used across open-weight models, and the rubric alone cannot act.

The empirical results span six benchmarks and seven models. The outcomes improve 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 authors describe this as exposing a meta-cognitive gap: unprompted models cannot reliably tell when their own context is decaying, but a lightweight rubric closes that gap. They frame when to compact as a capability scaffolds can supply without training.