STAIR: Improving Reasoning via Inter-query Token Reuse
September 29, 2026
STAIR (Stale-Token Attention for Inter-query Reuse) addresses how retained conversation history affects LLM accuracy by capturing keys and values from previous responses in a fixed bank. This allows models to redirect current queries to reuse computation from earlier problems in the same session.
HOW THIS AFFECTS YOU
●
builderYou can potentially improve reasoning accuracy in multi-turn agents by implementing efficient context reuse.
●
researcherThis provides a mechanism to manage the interference caused by stale tokens in long-context conversations.