SAKI: Score-Aware Low-Rank Key Indexing with Random-Matrix Noise Correction for KV Retrieval

Name: SAKI: Score-Aware Low-Rank Key Indexing with Random-Matrix Noise Correction for KV Retrieval
Uploaded: 2026-10-15T06:41:25.406628+00:00
Duration: 393 s
Description: SAKI develops an optimal rank-r linear score map to minimize distortion when compressing key information for retrieval. It proposes an asymmetric low-rank bilinear factorization that outperforms existing methods like key-PCA across all tested models and ranks, achieving significant recall improvements.

SAKI develops an optimal rank-r linear score map to minimize distortion when compressing key information for retrieval. It proposes an asymmetric low-rank bilinear factorization that outperforms existing methods like key-PCA across all tested models and ranks, achieving significant recall improvements.

Open the full episode