What's this about?
A paper (54 upvotes) letting models control their own attention: declaring `<global>`, `<focus>` or `<local>` inside the chain of thought, instead of scanning the whole KV cache every step.
What exactly is the KV cache?
The model's working memory of everything in context. Standard attention re-reads all of it per token — most of which is irrelevant most of the time. Declared modes let the model say "I only need this section now."
Why is this worth noting?
Long contexts are the cost center of agentic AI, and this makes attention scope a *decision the model makes* rather than fixed architecture — self-directed resource management, thematically kin to the week's self-evolution papers.