Language Models Can Control Their Own Attention
Declarative Attention lets a model state in its chain-of-thought where it needs to attend (global, focus, or local), and the inference engine treats those declarations like tool calls to skip most KV-cache reads. Zero-shot on 15 long-context tasks with Gemma-4-31B and Qwen-3.6-27B it cuts attended tokens by 52% and 31% with accuracy losses of 1.3-2.8pp that shrink with scale.
Why it matters
Turns KV-cache sparsification from an external-scoring problem into a model-declared one — O(1)-per-step attention control with off-the-shelf models, no training required.
Importance: 2/5
Notable paper: training-free KV-cache sparsification via model-declared attention scope