Let the model point to the tokens it actually needs
Long-context models may not need to scan their entire KV cache every time they answer. Declarative Attention has the model announce whether it needs the full context, a specific region, or only recent output, cutting attended tokens by 52.0% on Gemma-4-31B and 31.1% on Qwen-3.6-27B across 15 tasks, with accuracy drops of 1.27 and 2.75 percentage points. The tradeoff is promising, but training could make the model’s declarations safer and more decisive.