{"id":"arxiv:2609.02737","title":"Language Models Can Control Their Own Attention","date":"2026-09-02","section":"LLMs: architectures, context, training","quality":null,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/","source_url":"https://arxiv.org/abs/2609.02737","text_available":true,"authors":["Namgyu Ho","Huzama Ahmad","Woosung Koh","Se-Young Yun"],"outgoing":[{"source":"arxiv:2609.02737","target":"url:selfrag.github.io","evidence":"id","id":"url:selfrag.github.io","title":"Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f2a6b92a158b280a6699d4fda0ed474777be788abf9ed561ab7d61da1da57ad1/"}],"incoming":[],"text_parts":[{"start":0,"end":15634,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-1.html"},{"start":15634,"end":30954,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-2.html"},{"start":30954,"end":46015,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-3.html"},{"start":46015,"end":61444,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-4.html"},{"start":61444,"end":77286,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-5.html"},{"start":77286,"end":91934,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-6.html"}],"contents":[{"title":"Language Models Can Control Their Own Attention","offset":15,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-1.html#text"},{"title":"Abstract","offset":66,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-1.html#text"},{"title":"1 Introduction","offset":1474,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-1.html#text"},{"title":"Contribution.","offset":3685,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-1.html#text"},{"title":"2 Declarative Attention","offset":6934,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-1.html#text"},{"title":"2.1 Prompt overview","offset":10181,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-1.html#text"},{"title":"2.2 Context delivery","offset":11311,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-1.html#text"},{"title":"Context segmentation.","offset":12165,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-1.html#text"},{"title":"Segment formatting.","offset":13188,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-1.html#text"},{"title":"2.3 Decode-time interventions","offset":14166,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-1.html#text"},{"title":"Parsing and mode transitions.","offset":14201,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-1.html#text"},{"title":"Block-aligned mask construction.","offset":14803,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-1.html#text"},{"title":"Custom integration with vLLM.","offset":15598,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-1.html#text"},{"title":"Scope and tradeoffs.","offset":16242,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-2.html#text"},{"title":"3 When does DA pay off?","offset":16949,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-2.html#text"},{"title":"Overview.","offset":16977,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-2.html#text"},{"title":"FFN roofline wall-time.","offset":17753,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-2.html#text"},{"title":"Attention roofline wall-time.","offset":18339,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-2.html#text"},{"title":"Attention dominates FFN at scale.","offset":18869,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-2.html#text"},{"title":"4 Experimental settings","offset":19690,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-2.html#text"},{"title":"Models.","offset":19718,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-2.html#text"},{"title":"Evaluation datasets.","offset":20120,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-2.html#text"},{"title":"Baselines.","offset":21213,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-2.html#text"},{"title":"Evaluation.","offset":21905,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-2.html#text"},{"title":"Metrics.","offset":22343,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-2.html#text"},{"title":"Inference settings.","offset":22963,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-2.html#text"},{"title":"5 Results","offset":23437,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-2.html#text"},{"title":"5.1 Main results","offset":23451,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-2.html#text"},{"title":"Accuracy drops modestly on average.","offset":23473,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-2.html#text"},{"title":"Attention cost falls below vanilla.","offset":24121,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-2.html#text"},{"title":"The attention mask is the source of the efficiency gain.","offset":24968,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-2.html#text"},{"title":"5.2 Model capability scaling","offset":26182,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-2.html#text"},{"title":"Stronger models close the accuracy gap.","offset":26694,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-2.html#text"},{"title":"Token savings are roughly scale-independent.","offset":27804,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-2.html#text"},{"title":"5.3 Context-length scaling","offset":28636,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-2.html#text"},{"title":"Accuracy holds while token saving grows with context.","offset":28668,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-2.html#text"},{"title":"5.4 Efficiency results","offset":29589,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-2.html#text"},{"title":"Theoretical wall-clock time on optimized inference setups.","offset":29617,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-2.html#text"},{"title":"6 Analysis","offset":32132,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-3.html#text"},{"title":"6.1 Efficiency impact of DA modes","offset":32147,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-3.html#text"},{"title":"6.2 Protocol adherence","offset":33185,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-3.html#text"},{"title":"7 Conclusion","offset":34348,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-3.html#text"},{"title":"8 Discussion, Limitations, and Future Work","offset":35347,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-3.html#text"},{"title":"8.1 Relevance of DA in the present and future","offset":35394,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-3.html#text"},{"title":"Relevance of global attention costs in current and future architectures.","offset":36369,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-3.html#text"},{"title":"Relevance of DA in agentic systems.","offset":38018,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-3.html#text"},{"title":"8.2 Further potential of DA with post-training and in agentic settings","offset":38779,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-3.html#text"},{"title":"Zero-shot DA already demonstrates significant attention savings and positive scaling.","offset":38855,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-3.html#text"},{"title":"8.3 Synergy with modern efficient inference techniques","offset":41367,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-3.html#text"},{"title":"Lightweight scan sparse attention methods can alleviate costs in global mode.","offset":41427,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-3.html#text"},{"title":"Synergy with speculative decoding.","offset":42013,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-3.html#text"},{"title":"8.4 DA as system-2 sparse attention","offset":42586,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-3.html#text"},{"title":"Toward system-2 KV cache offloading.","offset":44093,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-3.html#text"},{"title":"Acknowledgements","offset":45495,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-3.html#text"},{"title":"Appendix A Related works","offset":45942,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-3.html#text"},{"title":"A.1 Dynamic sparse attention at decode","offset":45971,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-3.html#text"},{"title":"A.2 KV-cache eviction","offset":49228,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-4.html#text"},{"title":"A.3 Complementary KV-cache axes","offset":50354,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-4.html#text"},{"title":"A.4 Model-controlled inference and self-generated control tokens","offset":51478,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-4.html#text"},{"title":"A.5 Self-declared and steered attention","offset":53072,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-4.html#text"},{"title":"Appendix B DA vLLM integration","offset":55380,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-4.html#text"},{"title":"Always-attended regions.","offset":55772,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-4.html#text"},{"title":"Per-kernel patches.","offset":56491,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-4.html#text"},{"title":"Scope: global attention layers only.","offset":57123,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-4.html#text"},{"title":"Block alignment.","offset":57751,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-4.html#text"},{"title":"Effect on decode latency.","offset":58119,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-4.html#text"},{"title":"Appendix C Background on inference costs and derivation of roofline wall-time","offset":58428,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-4.html#text"},{"title":"C.1 Prefill vs decode","offset":58510,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-4.html#text"},{"title":"C.2 Decode cost components","offset":59027,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-4.html#text"},{"title":"FFN.","offset":59216,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-4.html#text"},{"title":"Attention.","offset":60012,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-4.html#text"},{"title":"C.3 The roofline at large batch","offset":60769,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-4.html#text"},{"title":"C.4 Roofline wall-time definition","offset":61405,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-4.html#text"},{"title":"C.5 Representative utilization values","offset":62228,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-5.html#text"},{"title":"MFU.","offset":62271,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-5.html#text"},{"title":"MBU.","offset":63244,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-5.html#text"},{"title":"C.6 The operating regime: production disaggregation at scale","offset":64276,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-5.html#text"},{"title":"C.7 Implication: attention dominates at scale","offset":66192,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-5.html#text"},{"title":"C.8 Worked decode wall-time decomposition for the headline models","offset":68670,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-5.html#text"},{"title":"Matmul (compute-bound).","offset":69395,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-5.html#text"},{"title":"Global memory (memory-bound).","offset":69948,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-5.html#text"},{"title":"Local memory (memory-bound).","offset":70664,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-5.html#text"},{"title":"C.9 KV bytes read per context token across current architectures","offset":73495,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-5.html#text"},{"title":"C.10 Attention share of decode wall-time at 1M-token contexts","offset":79397,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-6.html#text"},{"title":"Appendix D Experimental details","offset":83641,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-6.html#text"},{"title":"D.1 Evaluation datasets","offset":83677,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-6.html#text"},{"title":"D.2 Validation of local judge","offset":84635,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-6.html#text"},{"title":"D.3 Sampling parameters","offset":85608,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-6.html#text"},{"title":"D.4 Current limitations and future potential","offset":85909,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-6.html#text"},{"title":"Closing the gap.","offset":87621,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-6.html#text"},{"title":"Appendix E Additional experimental results","offset":88348,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-6.html#text"},{"title":"Appendix F Prompt construction","offset":88730,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-6.html#text"},{"title":"Context segmentation.","offset":88765,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-6.html#text"},{"title":"Magic-chunk delivery.","offset":90186,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-6.html#text"},{"title":"Instructions for reporting errors","offset":90970,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/part-6.html#text"}],"text_code_points":91934,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d92e296c1a145dd5abf7d3efa06faf9f01243de35f8f197fc1e6221cba9fe442/source.md"}