Language Models Can Control Thier Own Attention

Language models spend most of their attention on a small fraction of context, yet they still read the entire KV cache to find it. Ask about a detail in a long conversation and every token of the reply may require a full scan. Most sparse methods still pay an O(N) tax just to decide what matters.

Declarative Attention takes the opposite bet. The model already knows which parts are relevant. It declares the scope inside its chain of thought, in three modes. Global: look at everything to navigate. Focus: attend to a specific region. Local: use only recent output to synthesize. The inference engine parses those declarations like tool calls and skips most of the cache read.

On off-the-shelf models, total attended tokens fall by roughly a third to a half, with modest accuracy drops that shrink as models get larger. This is still zero-shot. The result is a lower bound, not a ceiling. Attention becomes a declared, reversible control signal the model can reason about in text, rather than a hidden, always-on scan.

This Ledger Entry expands how the reader thinks by treating long-context cost as a problem of self-declared scope, not of ever-larger caches or ever-heavier proxy scorers.