Language Models Can Control Their Own Attention
Language models spend most of their attention on a small fraction of context, yet they read the entire KV cache to find the few tokens that matter. If the user asks about a previous detail in a 1M-token conversation, global attention layers must scan the full context to generate each token of the reply. A prominent approach mitigates this cost by pre-selecting relevant tokens via lightweight proxy scores, but this extrinsic scoring still incurs O(N) per step. We take an intrinsic approach motivated by the simple question: wouldn't the model already know which parts of the context are relevant?
Lineage graph
Paper → model → repo connections mined from source citations (Tier-1 exact match).
Why these links exist
Every edge carries a method, confidence, and the source snippet that justified it — so bad links are debuggable.
- FuzzySimilar title/name (fuzzy) · 87%lllyasviel/ControlNet-v1-1 →
“Fuzzy title match (0.94): “Language Models Can Control Their Own Attention” ≈ “lllyasviel/ControlNet-v1-1””
- FuzzySimilar title/name (fuzzy) · 87%lllyasviel/ControlNet →
“Fuzzy title match (0.94): “Language Models Can Control Their Own Attention” ≈ “lllyasviel/ControlNet””
- PossiblePossibly related (embedding) · 55%Breakthrough in long-context efficiency announced →
- PossiblePossibly related (embedding) · 92%Language Models Can Control Their Own Attention [R] →
- LinkedLinked via arxiv author · 85%Namgyu Ho →
“Language Models Can Control Their Own Attention”
- LinkedLinked via arxiv author · 85%Huzama Ahmad →
“Language Models Can Control Their Own Attention”
- LinkedLinked via arxiv author · 85%Woosung Koh →
“Language Models Can Control Their Own Attention”
- LinkedLinked via arxiv author · 85%Se-Young Yun →
“Language Models Can Control Their Own Attention”