{"paper":{"title":"CompactAttention: Accelerating Chunked Prefill with Block-Union KV Selection","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Block-union KV selection builds minimal tables so chunked prefill attention runs up to 2.72 times faster while staying close to dense accuracy.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Beomseok Kang, Dongwon Jo, Jae-Joon Kim, Jiwon Song","submitted_at":"2026-05-16T06:47:41Z","abstract_excerpt":"Chunked prefill has become a widely adopted serving strategy for long-context large language models, but efficient attention computation in this regime remains challenging. Existing sparse attention methods are primarily designed for one-shot prefill and do not translate efficiently to chunked prefill: block-sparse kernels lose efficiency when the query length is limited by the chunk size, while fine-grained pattern search becomes costly when repeated over the accumulated KV cache at every chunk. QUOKA, a recent method that directly targets chunked prefill, avoids sparse-kernel overhead but re"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"On LLaMA-3.1-8B-Instruct, CompactAttention maintains accuracy close to dense attention on the RULER benchmark while delivering up to 2.72× attention speedup at 128K context length under chunked prefill.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The construction of GQA-aware per-group KV block tables via Q-block union and intra-group union produces the minimal tables that preserve all KV blocks selected by the input masks under paged execution constraints without missing query-specific entries.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"CompactAttention accelerates chunked-prefill attention via Block-Union KV Selection, delivering up to 2.72x speedup at 128K context on LLaMA-3.1-8B while matching dense accuracy on RULER.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Block-union KV selection builds minimal tables so chunked prefill attention runs up to 2.72 times faster while staying close to dense accuracy.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"2076819c7affe5a6e61cdc4bfb3bb202c05ee3a9d9a7879c895920213ea15900"},"source":{"id":"2605.16839","kind":"arxiv","version":1},"verdict":{"id":"52526f39-cdc7-4f46-b6e5-98baadde51ba","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T21:15:41.130579Z","strongest_claim":"On LLaMA-3.1-8B-Instruct, CompactAttention maintains accuracy close to dense attention on the RULER benchmark while delivering up to 2.72× attention speedup at 128K context length under chunked prefill.","one_line_summary":"CompactAttention accelerates chunked-prefill attention via Block-Union KV Selection, delivering up to 2.72x speedup at 128K context on LLaMA-3.1-8B while matching dense accuracy on RULER.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The construction of GQA-aware per-group KV block tables via Q-block union and intra-group union produces the minimal tables that preserve all KV blocks selected by the input masks under paged execution constraints without missing query-specific entries.","pith_extraction_headline":"Block-union KV selection builds minimal tables so chunked prefill attention runs up to 2.72 times faster while staying close to dense accuracy."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.16839/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_title_agreement","ran_at":"2026-05-19T21:31:19.236773Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T21:20:56.776000Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"claim_evidence","ran_at":"2026-05-19T18:41:56.323268Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T18:33:26.395742Z","status":"skipped","version":"1.0.0","findings_count":0}],"snapshot_sha256":"c456aa911eb1598e933d3ef7f12087df499110ce49fe6daf90c19f3d98342b88"},"references":{"count":28,"sample":[{"doi":"","year":2025,"title":"OpenAI GPT-5 System Card","work_id":"ca87689a-0d29-4476-b504-b65dbbb08af4","ref_index":1,"cited_arxiv_id":"2601.03267","is_internal_anchor":true},{"doi":"","year":2026,"title":"System card: Claude Opus 4.6","work_id":"d72a2098-05b3-4aae-ad94-f93a8fc5576d","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Gemini 3 Pro model card","work_id":"267390f5-3359-44fc-855f-79a2c06c7d30","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2026,"title":"Kimi K2.5: Visual Agentic Intelligence","work_id":"d690be8f-5d53-49b0-b1e7-79668eb8fcdb","ref_index":4,"cited_arxiv_id":"2602.02276","is_internal_anchor":true},{"doi":"","year":2026,"title":"Deepseek-v4: Towards highly efficient million-token context intelligence","work_id":"b5a5843d-47c0-432e-88d5-c9eaac555a7f","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":28,"snapshot_sha256":"1e70316cabd47fd7bd15ab7a67cabec1e4f42f290186508b9ff28f8cd85a2a26","internal_anchors":9},"formal_canon":{"evidence_count":1,"snapshot_sha256":"13fb3fec9d7324a45aef6c07420d0047a3ef2724a78ed97d2733c55fc98a9517"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}