AI Post Transformers · Episode Companion
KV Cache Eviction Through an Information Bottleneck Lens
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.