One email each morning — yesterday's AI, sortedGet it in your inbox

METAL LAB

Fractional Decay KV-Cache: Ownership-Aware Memory Management for Improved Inference Relevancy in Dialog Systems

arXiv:2608.180982026-08-20

Key-value (KV) caching is essential for efficient autoregressive inference in transformer based dialog systems, yet existing strategies treat all cached entries uniformly or apply coarse eviction heuristics that fail to adapt as dialog topics evolve. We propose Fractional Decay KV-Cache (FD-KVC), a novel algorithm that maintains a dual-channel scoring mechanism for each cached KV pair: a cumulative attention channel that tracks aggregate importance

Authors · Sukanta Ganguly

Read on arXiv

Latest papers

All papers →

Latest from METAL LAB