4 sink tokens + 64-token window matches distilled linear attention, no training needed

burny_tech · x · 2026-09-14

The author reports that while millions have been spent post-training LLMs into linear attention to remove the KV cache bottleneck, a training-free trick—keeping just 4 initial tokens plus a 64-token sliding window—matches or beats distilled linear models on almost every benchmark.

The finding suggests much of the benefit of linearized models may come simply from preserving initial tokens (an attention-sink-like effect) rather than from the distillation itself, hinting at far cheaper alternatives to some KV-cache optimization work. Details in the thread.

Original post →

More from Infra

Infra channel →