After a Document Edit, Recompute a Contiguous Window Around It, Not the Attention-Important Positions

arXiv 2609.17983·medium signalXBlueskyLinkedInCopy link

arXiv 2609.17983 (16 Sep 2026) formulates stale KV-cache repair as budgeted recomputation and compares training-free position-selection policies on a factual RAG benchmark with matched direct and derived edits. All policies handle direct edits; derived edits separate them, and at the primary budget a contiguous edit-local window recovers at least 0.94 of the post-edit answer margin, substantially beating attention-based, KV-deviation and structural selectors. The mechanistic reason is that scattered positions inherit surrounding staleness even when they look important under clean-state transplantation, and the edit-local advantage depends on adjacency, largely disappearing when the answer-bearing text sits downstream.