{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2026:64TM67QCYXLMFOU2XSEVWSQW2P","short_pith_number":"pith:64TM67QC","canonical_record":{"source":{"id":"2602.03216","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-02-03T07:31:14Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"2b9442986cf3fbf8890f25bbcbaca7bb9db8b35d35b77588071f66f96f28a229","abstract_canon_sha256":"d43213bf51fc6570195b71c7bcbf12e6913b9d126c557005b7c5ff6af3d848ed"},"schema_version":"1.0"},"canonical_sha256":"f726cf7e02c5d6c2ba9abc895b4a16d3f0119dcdac45872a55c2ba8e55f2e7e0","source":{"kind":"arxiv","id":"2602.03216","version":3},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2602.03216","created_at":"2026-06-01T01:02:31Z"},{"alias_kind":"arxiv_version","alias_value":"2602.03216v3","created_at":"2026-06-01T01:02:31Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2602.03216","created_at":"2026-06-01T01:02:31Z"},{"alias_kind":"pith_short_12","alias_value":"64TM67QCYXLM","created_at":"2026-06-01T01:02:31Z"},{"alias_kind":"pith_short_16","alias_value":"64TM67QCYXLMFOU2","created_at":"2026-06-01T01:02:31Z"},{"alias_kind":"pith_short_8","alias_value":"64TM67QC","created_at":"2026-06-01T01:02:31Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2026:64TM67QCYXLMFOU2XSEVWSQW2P","target":"record","payload":{"canonical_record":{"source":{"id":"2602.03216","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-02-03T07:31:14Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"2b9442986cf3fbf8890f25bbcbaca7bb9db8b35d35b77588071f66f96f28a229","abstract_canon_sha256":"d43213bf51fc6570195b71c7bcbf12e6913b9d126c557005b7c5ff6af3d848ed"},"schema_version":"1.0"},"canonical_sha256":"f726cf7e02c5d6c2ba9abc895b4a16d3f0119dcdac45872a55c2ba8e55f2e7e0","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-01T01:02:31.952272Z","signature_b64":"TLyfmCBL2VF/KHiGW++VGpWzJHWk2uW+cHGEOlpbP88KH9G0jceRSkn+pQRXoGt9lu/ZiXmU+jOyYLGPjEk2BA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"f726cf7e02c5d6c2ba9abc895b4a16d3f0119dcdac45872a55c2ba8e55f2e7e0","last_reissued_at":"2026-06-01T01:02:31.951364Z","signature_status":"signed_v1","first_computed_at":"2026-06-01T01:02:31.951364Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2602.03216","source_version":3,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-06-01T01:02:31Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"jPAH3eYYYsMAF89NTj1w3YitxwN4PG9J3iA0xo1iE7rV4bVXJj+EknvaC3Vs3t0FziEXfdUN+3DFt5ZncM2ICw==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-07-18T19:05:10.503384Z"},"content_sha256":"7c5a94b73e6074dc3299fad0b75e4d73641540dc7f72dc595fd418ad89a1e01a","schema_version":"1.0","event_id":"sha256:7c5a94b73e6074dc3299fad0b75e4d73641540dc7f72dc595fd418ad89a1e01a"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2026:64TM67QCYXLMFOU2XSEVWSQW2P","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"Token Sparse Attention: Efficient Long-Context Inference with Interleaved Token Selection","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Dynamic per-head token selection during attention compresses computation then restores full sequence length to speed up long-context inference.","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Beomseok Kang, Dongwon Jo, Jae-Joon Kim, Jiwon Song","submitted_at":"2026-02-03T07:31:14Z","abstract_excerpt":"The quadratic complexity of attention remains the central bottleneck in long-context inference for large language models. Prior acceleration methods either sparsify the attention map with structured patterns or permanently evict tokens at specific layers, which can retain irrelevant tokens or rely on irreversible early decisions despite the layer-/head-wise dynamics of token importance. In this paper, we propose Token Sparse Attention, a lightweight and dynamic token-level sparsification mechanism that compresses per-head $Q$, $K$, $V$ to a reduced token set during attention and then decompres"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Experimental results show that Token Sparse Attention consistently improves accuracy-latency trade-off, achieving up to ×3.23 attention speedup at 128K context with less than 1% accuracy degradation.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That dynamic per-head token selection followed by decompression reliably preserves necessary information so that tokens can be meaningfully reconsidered in subsequent layers without irreversible loss.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Token Sparse Attention uses dynamic per-head token compression and decompression during attention to achieve up to 3.23x speedup at 128K context with under 1% accuracy loss.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Dynamic per-head token selection during attention compresses computation then restores full sequence length to speed up long-context inference.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"fb6d2138917f42c20afa03849d99cc773ff8f83688760f2ead3fe9df57ac9a3a"},"source":{"id":"2602.03216","kind":"arxiv","version":3},"verdict":{"id":"bd905358-39ae-4fed-8bfc-b21cff4cece5","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T08:28:07.868396Z","strongest_claim":"Experimental results show that Token Sparse Attention consistently improves accuracy-latency trade-off, achieving up to ×3.23 attention speedup at 128K context with less than 1% accuracy degradation.","one_line_summary":"Token Sparse Attention uses dynamic per-head token compression and decompression during attention to achieve up to 3.23x speedup at 128K context with under 1% accuracy loss.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That dynamic per-head token selection followed by decompression reliably preserves necessary information so that tokens can be meaningfully reconsidered in subsequent layers without irreversible loss.","pith_extraction_headline":"Dynamic per-head token selection during attention compresses computation then restores full sequence length to speed up long-context inference."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2602.03216/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"verdict_id":"bd905358-39ae-4fed-8bfc-b21cff4cece5"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-06-01T01:02:31Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"5qlkgwH2iLVOr7SIHR7LHRTDfuNDKv12GjFVfRj6lsj8QdclzU3RQx5ETUTK6RXBPNXBauy1FGm2uHFccdCvCg==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-07-18T19:05:10.503841Z"},"content_sha256":"8b305adbf25cd1b7441687489e0bb5ce3aa6cfd690f2946707a262a9485acdeb","schema_version":"1.0","event_id":"sha256:8b305adbf25cd1b7441687489e0bb5ce3aa6cfd690f2946707a262a9485acdeb"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/64TM67QCYXLMFOU2XSEVWSQW2P/bundle.json","state_url":"https://pith.science/pith/64TM67QCYXLMFOU2XSEVWSQW2P/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/64TM67QCYXLMFOU2XSEVWSQW2P/bundle.json","status":"primary"}],"public_keys":[{"key_id":"pith-v1-2026-05","algorithm":"ed25519","format":"raw","public_key_b64":"stVStoiQhXFxp4s2pdzPNoqVNBMojDU/fJ2db5S3CbM=","public_key_hex":"b2d552b68890857171a78b36a5dccf368a953413288c353f7c9d9d6f94b709b3","fingerprint_sha256_b32_first128bits":"RVFV5Z2OI2J3ZUO7ERDEBCYNKS","fingerprint_sha256_hex":"8d4b5ee74e4693bcd1df2446408b0d54","rotates_at":null,"url":"https://pith.science/pith-signing-key.json","notes":"Pith uses this Ed25519 key to sign canonical record SHA-256 digests. Verify with: ed25519_verify(public_key, message=canonical_sha256_bytes, signature=base64decode(signature_b64))."}],"merge_version":"pith-open-graph-merge-v1","built_at":"2026-07-18T19:05:10Z","links":{"resolver":"https://pith.science/pith/64TM67QCYXLMFOU2XSEVWSQW2P","bundle":"https://pith.science/pith/64TM67QCYXLMFOU2XSEVWSQW2P/bundle.json","state":"https://pith.science/pith/64TM67QCYXLMFOU2XSEVWSQW2P/state.json","well_known_bundle":"https://pith.science/.well-known/pith/64TM67QCYXLMFOU2XSEVWSQW2P/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:64TM67QCYXLMFOU2XSEVWSQW2P","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"d43213bf51fc6570195b71c7bcbf12e6913b9d126c557005b7c5ff6af3d848ed","cross_cats_sorted":["cs.LG"],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-02-03T07:31:14Z","title_canon_sha256":"2b9442986cf3fbf8890f25bbcbaca7bb9db8b35d35b77588071f66f96f28a229"},"schema_version":"1.0","source":{"id":"2602.03216","kind":"arxiv","version":3}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2602.03216","created_at":"2026-06-01T01:02:31Z"},{"alias_kind":"arxiv_version","alias_value":"2602.03216v3","created_at":"2026-06-01T01:02:31Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2602.03216","created_at":"2026-06-01T01:02:31Z"},{"alias_kind":"pith_short_12","alias_value":"64TM67QCYXLM","created_at":"2026-06-01T01:02:31Z"},{"alias_kind":"pith_short_16","alias_value":"64TM67QCYXLMFOU2","created_at":"2026-06-01T01:02:31Z"},{"alias_kind":"pith_short_8","alias_value":"64TM67QC","created_at":"2026-06-01T01:02:31Z"}],"graph_snapshots":[{"event_id":"sha256:8b305adbf25cd1b7441687489e0bb5ce3aa6cfd690f2946707a262a9485acdeb","target":"graph","created_at":"2026-06-01T01:02:31Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","text":"Experimental results show that Token Sparse Attention consistently improves accuracy-latency trade-off, achieving up to ×3.23 attention speedup at 128K context with less than 1% accuracy degradation."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That dynamic per-head token selection followed by decompression reliably preserves necessary information so that tokens can be meaningfully reconsidered in subsequent layers without irreversible loss."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"Token Sparse Attention uses dynamic per-head token compression and decompression during attention to achieve up to 3.23x speedup at 128K context with under 1% accuracy loss."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Dynamic per-head token selection during attention compresses computation then restores full sequence length to speed up long-context inference."}],"snapshot_sha256":"fb6d2138917f42c20afa03849d99cc773ff8f83688760f2ead3fe9df57ac9a3a"},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"integrity":{"available":true,"clean":true,"detectors_run":[],"endpoint":"/pith/2602.03216/integrity.json","findings":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938","summary":{"advisory":0,"by_detector":{},"critical":0,"informational":0}},"paper":{"abstract_excerpt":"The quadratic complexity of attention remains the central bottleneck in long-context inference for large language models. Prior acceleration methods either sparsify the attention map with structured patterns or permanently evict tokens at specific layers, which can retain irrelevant tokens or rely on irreversible early decisions despite the layer-/head-wise dynamics of token importance. In this paper, we propose Token Sparse Attention, a lightweight and dynamic token-level sparsification mechanism that compresses per-head $Q$, $K$, $V$ to a reduced token set during attention and then decompres","authors_text":"Beomseok Kang, Dongwon Jo, Jae-Joon Kim, Jiwon Song","cross_cats":["cs.LG"],"headline":"Dynamic per-head token selection during attention compresses computation then restores full sequence length to speed up long-context inference.","license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-02-03T07:31:14Z","title":"Token Sparse Attention: Efficient Long-Context Inference with Interleaved Token Selection"},"references":{"count":0,"internal_anchors":0,"resolved_work":0,"sample":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2602.03216","kind":"arxiv","version":3},"verdict":{"created_at":"2026-05-16T08:28:07.868396Z","id":"bd905358-39ae-4fed-8bfc-b21cff4cece5","model_set":{"reader":"grok-4.3"},"one_line_summary":"Token Sparse Attention uses dynamic per-head token compression and decompression during attention to achieve up to 3.23x speedup at 128K context with under 1% accuracy loss.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"Dynamic per-head token selection during attention compresses computation then restores full sequence length to speed up long-context inference.","strongest_claim":"Experimental results show that Token Sparse Attention consistently improves accuracy-latency trade-off, achieving up to ×3.23 attention speedup at 128K context with less than 1% accuracy degradation.","weakest_assumption":"That dynamic per-head token selection followed by decompression reliably preserves necessary information so that tokens can be meaningfully reconsidered in subsequent layers without irreversible loss."}},"verdict_id":"bd905358-39ae-4fed-8bfc-b21cff4cece5"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:7c5a94b73e6074dc3299fad0b75e4d73641540dc7f72dc595fd418ad89a1e01a","target":"record","created_at":"2026-06-01T01:02:31Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"d43213bf51fc6570195b71c7bcbf12e6913b9d126c557005b7c5ff6af3d848ed","cross_cats_sorted":["cs.LG"],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-02-03T07:31:14Z","title_canon_sha256":"2b9442986cf3fbf8890f25bbcbaca7bb9db8b35d35b77588071f66f96f28a229"},"schema_version":"1.0","source":{"id":"2602.03216","kind":"arxiv","version":3}},"canonical_sha256":"f726cf7e02c5d6c2ba9abc895b4a16d3f0119dcdac45872a55c2ba8e55f2e7e0","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"f726cf7e02c5d6c2ba9abc895b4a16d3f0119dcdac45872a55c2ba8e55f2e7e0","first_computed_at":"2026-06-01T01:02:31.951364Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-06-01T01:02:31.951364Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"TLyfmCBL2VF/KHiGW++VGpWzJHWk2uW+cHGEOlpbP88KH9G0jceRSkn+pQRXoGt9lu/ZiXmU+jOyYLGPjEk2BA==","signature_status":"signed_v1","signed_at":"2026-06-01T01:02:31.952272Z","signed_message":"canonical_sha256_bytes"},"source_id":"2602.03216","source_kind":"arxiv","source_version":3}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:7c5a94b73e6074dc3299fad0b75e4d73641540dc7f72dc595fd418ad89a1e01a","sha256:8b305adbf25cd1b7441687489e0bb5ce3aa6cfd690f2946707a262a9485acdeb"],"state_sha256":"5948ee2f5da90eb058de483914fae9b68db148fd8e2a71f9b8eca50498c48c81"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"UdzDziVepcLeEcWt56b1sYhQY3EAqCG52fNvKaS1e1x571X4jYKu64JEVS1zuUWz4lUvwk9KyTgKMEoaOzPMCQ==","signed_message":"bundle_sha256_bytes","signed_at":"2026-07-18T19:05:10.506507Z","bundle_sha256":"f08d62ea6b5804d07b8900e1eda5bab5c0e7473efa2d4b08a6931435be142d80"}}