Declarative Attention: LLMs Slash KV Cache Reads 31-52% Zero-Shot

chaumian · x · 2026-09-04

The arXiv paper "Language Models Can Control Their Own Attention" introduces Declarative Attention (DA), a protocol where the model itself declares where to attend in its chain-of-thought via <global>/<focus>/<local> modes, letting the inference engine skip most KV cache reads. Zero-shot across 15 long-context tasks, DA cuts total attended tokens during decoding by 52.0% (Gemma-4-31B) and 31.1% (Qwen-3.6-27B) with accuracy drops of only 1.27pp/2.75pp that shrink with scale — a new axis of sparse attention.

Related event: Paper Lets Language Models Control Their Own Attention, Cutting Decode Cost 52%(3 posts)→

Original post →

More from Research

Research channel →