An Honest Negative Result on KV-Cache Eviction: Fixed-Lag 'Demonstrated Utility' Collapses Onto Plain Attention on Real Benchmarks

arXiv 2607.24667·medium signalXBlueskyLinkedInCopy link

'Eviction as Estimation' (arXiv 2607.24667, July 27) recasts KV-cache eviction as estimation of a hidden reuse signal along a commit-lag axis, with StreamingLLM/H2O/SnapKV at lag H=0 and Belady's optimum at full future knowledge, then fills the missing middle: wait a bounded number of steps, observe which items a correct near-future prediction actually attended to, and only then evict. The training-free policy RMM strictly generalizes H2O. In controlled settings where reuse is endogenous and time-separated it identifies used memory far better than accumulated attention — but run inside NVIDIA's KVPress harness against KVPress's own SnapKV, H2O, and StreamingLLM implementations, the advantage largely vanishes: RMM ties H2O on single-turn QA and loses to both in streaming multi-turn. The authors state plainly that the contribution is the framework and 'an honest map of when measuring beats accumulating, not a new state of the art.'