The Key-Value (KV) cache is the main memory bottleneck of long-context LLM inference. Token eviction methods such as SnapKV, H2O, AdaKV, and L2Norm shrink it, but they give every transformer layer the same budget. The MOSAIC paper argues this uniform allocation is a limitation; the abstract starts to explain why, noting that layers differ, though the sentence is cut off in the source.
MOSAIC stands for Multi-Objective Search for Allocating Inference Cache Across LLM Decoder Layers. Instead of a fixed global budget, it searches for a per-layer allocation. The abstract does not detail the objectives or search method, but the name and framing indicate a multi-objective optimization approach.
The source is a single arXiv abstract, so no experimental results are available. The contribution is the problem formulation: moving from a one-size-fits-all eviction budget to a per-layer, multi-objective allocation.