BWhat changed
Reused selected token indices across layers to reduce sparse-attention indexing cost, with calibration-based layer selection and an optional multi-layer distillation objective.
Scientific scope
Reused selected token indices across layers to reduce sparse-attention indexing cost, with calibration-based layer selection and an optional multi-layer distillation objective. DSA token selection is inherited. Approximate cross-layer index reuse is not exact equality to every layer's original selection. Reported speed/quality tradeoffs depend on calibration, retained indexers, model and hardware; later GLM naming is not another invention.
Evidence
The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. DSA token selection is inherited. Approximate cross-layer index reuse is not exact equality to every layer's original selection. Reported speed/quality tradeoffs depend on calibration, retained indexers, model and hardware; later GLM naming is not another invention.
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.
- IndexCache: Accelerating Sparse Attention via Cross-Layer Index Reuse
- Bai et al.: IndexCache cross-layer sparse-index reuse — official method source
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