{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:6BBRQ4MA57NCEYPZFIT5URYVFN","short_pith_number":"pith:6BBRQ4MA","schema_version":"1.0","canonical_sha256":"f043187180efda2261f92a27da47152b5d8db587eb37659aa4128dea346384d3","source":{"kind":"arxiv","id":"2412.19255","version":2},"attestation_state":"computed","paper":{"title":"Multi-matrix Factorization Attention","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Daxin Jiang, Heung-Yeung Shum, Houyi Li, Jingcheng Hu, Shuigeng Zhou, Xiangyu Zhang, Yinmin Zhang, Zili Wang","submitted_at":"2024-12-26T15:45:45Z","abstract_excerpt":"We propose novel attention architectures, Multi-matrix Factorization Attention (MFA) and MFA-Key-Reuse (MFA-KR). Existing variants for standard Multi-Head Attention (MHA), including SOTA methods like MLA, fail to maintain as strong performance under stringent Key-Value cache (KV cache) constraints. MFA enhances model capacity by efficiently scaling up both the number and dimension of attention heads through low-rank matrix factorization in the Query-Key (QK) circuit. Extending MFA, MFA-KR further reduces memory requirements by repurposing the key cache as value through value projection re-para"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2412.19255","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2024-12-26T15:45:45Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"c9e950ba506f4bfa4860ad3fdb6bb991894bf96bc243ef8fd60da5d4f4731d0c","abstract_canon_sha256":"0b797382e7a5887c7f8619b7bc6510d9acd6060925ac6822193e4c718744be69"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:00:45.358304Z","signature_b64":"GnDetOsmO+ddcMFUfH5abe15KH4m8mLdqJyZUKfJ+OmErUp9RWXVjUxVAeaCc1ciCBkamfdqtRPDZkahoujQBQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"f043187180efda2261f92a27da47152b5d8db587eb37659aa4128dea346384d3","last_reissued_at":"2026-07-05T10:00:45.357737Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:00:45.357737Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Multi-matrix Factorization Attention","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Daxin Jiang, Heung-Yeung Shum, Houyi Li, Jingcheng Hu, Shuigeng Zhou, Xiangyu Zhang, Yinmin Zhang, Zili Wang","submitted_at":"2024-12-26T15:45:45Z","abstract_excerpt":"We propose novel attention architectures, Multi-matrix Factorization Attention (MFA) and MFA-Key-Reuse (MFA-KR). Existing variants for standard Multi-Head Attention (MHA), including SOTA methods like MLA, fail to maintain as strong performance under stringent Key-Value cache (KV cache) constraints. MFA enhances model capacity by efficiently scaling up both the number and dimension of attention heads through low-rank matrix factorization in the Query-Key (QK) circuit. Extending MFA, MFA-KR further reduces memory requirements by repurposing the key cache as value through value projection re-para"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2412.19255","kind":"arxiv","version":2},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2412.19255/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"},"aliases":[{"alias_kind":"arxiv","alias_value":"2412.19255","created_at":"2026-07-05T10:00:45.357798+00:00"},{"alias_kind":"arxiv_version","alias_value":"2412.19255v2","created_at":"2026-07-05T10:00:45.357798+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2412.19255","created_at":"2026-07-05T10:00:45.357798+00:00"},{"alias_kind":"pith_short_12","alias_value":"6BBRQ4MA57NC","created_at":"2026-07-05T10:00:45.357798+00:00"},{"alias_kind":"pith_short_16","alias_value":"6BBRQ4MA57NCEYPZ","created_at":"2026-07-05T10:00:45.357798+00:00"},{"alias_kind":"pith_short_8","alias_value":"6BBRQ4MA","created_at":"2026-07-05T10:00:45.357798+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2505.21487","citing_title":"Hardware-Efficient Attention for Fast Decoding","ref_index":31,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/6BBRQ4MA57NCEYPZFIT5URYVFN","json":"https://pith.science/pith/6BBRQ4MA57NCEYPZFIT5URYVFN.json","graph_json":"https://pith.science/api/pith-number/6BBRQ4MA57NCEYPZFIT5URYVFN/graph.json","events_json":"https://pith.science/api/pith-number/6BBRQ4MA57NCEYPZFIT5URYVFN/events.json","paper":"https://pith.science/paper/6BBRQ4MA"},"agent_actions":{"view_html":"https://pith.science/pith/6BBRQ4MA57NCEYPZFIT5URYVFN","download_json":"https://pith.science/pith/6BBRQ4MA57NCEYPZFIT5URYVFN.json","view_paper":"https://pith.science/paper/6BBRQ4MA","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2412.19255&json=true","fetch_graph":"https://pith.science/api/pith-number/6BBRQ4MA57NCEYPZFIT5URYVFN/graph.json","fetch_events":"https://pith.science/api/pith-number/6BBRQ4MA57NCEYPZFIT5URYVFN/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/6BBRQ4MA57NCEYPZFIT5URYVFN/action/timestamp_anchor","attest_storage":"https://pith.science/pith/6BBRQ4MA57NCEYPZFIT5URYVFN/action/storage_attestation","attest_author":"https://pith.science/pith/6BBRQ4MA57NCEYPZFIT5URYVFN/action/author_attestation","sign_citation":"https://pith.science/pith/6BBRQ4MA57NCEYPZFIT5URYVFN/action/citation_signature","submit_replication":"https://pith.science/pith/6BBRQ4MA57NCEYPZFIT5URYVFN/action/replication_record"}},"created_at":"2026-07-05T10:00:45.357798+00:00","updated_at":"2026-07-05T10:00:45.357798+00:00"}