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).
More from Infra
- Three myths of hosted LLM inference: sticker prices, interchangeable endpoints, and self-hosting — TangeloOk9486 · 2026-09-27
- Laptop engine streams a 35B model from SSD at 9.4 tok/s, beating GPT-OSS 20B — ImBadGuyInEveryStory · 2026-09-27
- World's fastest panel QR factorization on B200: how a GPU MODE contestant cracked chained dependencies — A_K_Nain · 2026-09-27
- TensorSharp open-source engine adds mixed document/image/video/audio inputs per request — fuzhongkai · 2026-09-27
- Pro-data center rally clashes with protesters as scholar defends AI infrastructure — neil_chilson · 2026-09-27
- Kafka in production: partition skew and rebalancing pauses bite hard — goyalshaliniuk · 2026-09-27