The abstract for PeekKV, posted on arXiv, opens by noting that the KV cache grows linearly with context length, making it a major memory bottleneck in LLM serving. It then points to recent progress in identifying the functional roles of attention heads, which has motivated a growing body of work on compression. The provided abstract is cut off mid-sentence at "compresses…", so the full method and results are not available in the source.
What is clear from the title and the opening is that PeekKV proposes a simple selection rule: look at the first content token to determine which attention heads matter, and keep only those in the cache. This would avoid storing the full KV cache for every head. The abstract does not yet provide details on how this selection is implemented or what savings it achieves, and no comparison to other compression methods is given in the excerpt.