Random Attention: Salesforce and UIUC find random KV cache eviction rivals handcrafted signals

jiqizhixin · x · 2026-09-27

Reasoning models generate thousands of tokens, and KV cache grows linearly with sequence length — for long-reasoning serving it quickly becomes the real system bottleneck. The standard answer is KV cache eviction with a fixed budget and an importance signal: accumulated attention, recent queries, redundancy tracking, value magnitudes or key statistics.

Salesforce AI Research and UIUC ask a near "anti-algorithmic" question: are the carefully designed selection signals themselves the problem?

Their answer is Random Attention — evict at random and drop the expensive scoring pass entirely. The study finds this surprisingly competitive (thread truncated; full results in the paper).

Original post →

More from Infra

Infra channel →