Researchers introduce Declarative Attention (DA), a protocol that has a language model explicitly declare in its chain-of-thought which parts of a long context it needs to attend to, letting the inference engine skip most of the KV cache read rather than scanning the full context for every token. Tested zero-shot across 15 long-context tasks on Gemma-4-31B and Qwen-3.6-27B, DA cuts total attended tokens during decoding by 31-52% with accuracy drops of only 1.3-2.75 percentage points that shrink as model scale increases.
