High-Fidelity KV Cache Summarization Using Entropy and Low-Rank Reconstruction

Interesting Approach. Curious about the latency tradeoff: OLS + SVD are much heavier than Top-K.Have you benchmarked end-to-end inference latency?

[dead]