Language Models Can Control Their Own Attention

Authors/date: Namgyu Ho and five co-authors; arXiv preprint; 2 September 2026. Contribution: Declarative Attention lets a model signal global, focused or local context regions so the inference engine can skip much of the KV-cache read. Authors report 31.1–52.0% fewer attended tokens with modest accuracy reductions on two off-the-shelf models. Limitations: zero-shot experiments, two model families and no production latency study. Paper