AI Post Transformers · Episode Companion

KV Cache Eviction Through an Information Bottleneck Lens

↗ arXiv:2604.25975 Yang, Tang, Zhen, Lv · 2026 Sichuan University · A*STAR IHPC Method: CAPKV

A first-principles attempt to unify attention-pattern eviction (SnapKV, H2O) and structure-aware eviction (KeyDiff, Knorm) under one Information Bottleneck objective — treating the surviving KV cache as a lossy compression of history that must stay maximally informative about a future query it hasn't seen yet.

References

Cited across the episode. Only the source paper has a resolvable arXiv ID; the rest link to scholar.google.com as provided.