H2O: recent/heavy-hitter KV cache eviction

What changed

Combined recency and accumulated attention heavy hitters in a dynamic KV-cache eviction policy and evaluated its memory/inference tradeoff.

Scientific scope

Combined recency and accumulated attention heavy hitters in a dynamic KV-cache eviction policy and evaluated its memory/inference tradeoff. Inherits cache eviction and sparse attention. Conditional submodular analysis is source-reported and not proof-audited here; model assumptions are not automatic neural-network facts. Token eviction is lossy and runtime gains depend on batch/workload/baseline.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Inherits cache eviction and sparse attention. Conditional submodular analysis is source-reported and not proof-audited here; model assumptions are not automatic neural-network facts. Token eviction is lossy and runtime gains depend on batch/workload/baseline.

Contributors

• Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.

• H$_2$O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models

Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction