PyramidKV: layer-dependent KV cache budgets

What changed

Allocated different KV-cache budgets across layers based on observed information-funneling patterns, retaining more entries in lower layers.

Scientific scope

Allocated different KV-cache budgets across layers based on observed information-funneling patterns, retaining more entries in lower layers. Inherits attention-based token selection and cache compression. Pyramidal behavior is empirical, not universal for every model/layer/task. Cache ratios and quality retention are setup-specific; deleted information is not losslessly compressed.

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 attention-based token selection and cache compression. Pyramidal behavior is empirical, not universal for every model/layer/task. Cache ratios and quality retention are setup-specific; deleted information is not losslessly compressed.

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.

• PyramidKV: Dynamic KV Cache Compression based on Pyramidal Information Funneling

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