Declarative Attention (DA)

Declarative Attention (DA) is an inference protocol introduced by Ho et al. (KAIST AI and Google DeepMind) in 2026. Rather than having the inference engine read the entire KV cache on every generated token, DA lets the model declare which parts of the context it is attending to, using a syntax embedded in the chain-of-thought. The inference engine parses the declarations and skips the rest of the KV cache read.

DA partitions generation into three modes. In global mode, the model reads the full context, used for navigation across the prompt. In focus mode, the model reads only a specific named region (a "magic chunk"), used for reasoning about a particular piece of context. In local mode, the model reads only the recent output, used for responding without re-reading the prompt. The model alternates between these modes freely during a single response.

On long-context tasks, DA reduces attended tokens during decoding by 52% on Gemma-4-31B and 31% on Qwen-3.6-27B, with accuracy drops of 1.27 percentage points and 2.75 percentage points respectively. The accuracy cost shrinks with model scale. The protocol runs on off-the-shelf models without retraining, requiring only a declarative attention syntax in the system prompt and an inference engine that respects the declarations.

The motivation is intrinsic rather than extrinsic. Instead of pre-selecting relevant tokens via a learned proxy scorer, the protocol trusts the model to know which parts of its own context are relevant. DA is related to but distinct from sparse attention, which selects tokens to attend to via learned or heuristic mechanisms rather than the model's own declarations.