Attention-guided trajectory compression beat SWE-bench baselines while cutting tokens 21.6% and cost 33.6%

arXiv 2609.08318·high signalXBlueskyLinkedInCopy link

AttnCompress segments an agent trajectory at perplexity spikes to keep code and log syntax intact, uses proxy attention weights to score how relevant each historical block is to the agent's current reasoning, and runs a dynamic rolling window that can recall context it previously dropped as the task evolves. On SWE-Bench-Verified and Multi-SWE-Bench it reached a 53.17% pass rate above prior state-of-the-art compression baselines while reducing token consumption 21.6% and total cost 33.6%, and the approach is model-agnostic across languages. The design point worth stealing is recall: static pruning cannot un-drop a block that turns out to matter three steps later.