Entropy, OLS, SVD tame KV cache spikes

Promising direction, but it is still proving a narrow claim: better reconstruction on a synthetic setup. The hard part is whether the extra math stays worth it once you measure real latency, throughput, and model-task variability.

• –Top-K’s failure mode here is plausible: most tokens look fine, but a few structurally important ones blow up error
• –Entropy is a smarter selection signal than raw magnitude, but it may be brittle when attention is diffuse for reasons that still matter downstream
• –OLS plus SVD shifts the bottleneck from memory to compute, so kernel efficiency and amortization cadence decide whether this is practical
• –The memory win is interesting, but the fairness assumptions matter; “lower memory” is only useful if end-to-end serving cost also improves
• –If it generalizes, this looks more like a long-context serving strategy than a simple pruning replacement