{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:JTCELKEJLJHS3FEFVTRRK34N25","short_pith_number":"pith:JTCELKEJ","schema_version":"1.0","canonical_sha256":"4cc445a8895a4f2d9485ace3156f8dd7750db1929cbd9e2ea3088063b8d1f9b0","source":{"kind":"arxiv","id":"2507.09394","version":1},"attestation_state":"computed","paper":{"title":"A Random Matrix Theory Perspective on the Learning Dynamics of Multi-head Latent Attention","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Brandon Reagen, Nandan Kumar Jha","submitted_at":"2025-07-12T20:31:07Z","abstract_excerpt":"In this work, we study how multi-head latent attention (MLA), a popular strategy for compressing key/value memory, affects a transformer's internal capacity during pretraining. Using a lightweight suite of Marchenko-Pastur (MP) diagnostics, we analyze the spectrum of the $W_{Q}W_{K}^\\top$ gram matrix throughout training, comparing three variants: the standard multi-head attention (MHA) baseline, MLA-PreRoPE with rotary applied before compression, and MLA-Decoupled, which shares a single rotary sub-vector across all heads. Our random matrix analysis reveals \\textbf{three key findings:} \\textbf{"},"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":"2507.09394","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-07-12T20:31:07Z","cross_cats_sorted":[],"title_canon_sha256":"c34a4357fd1b40678bdbed2fc75d7463434435f91b4fb1900c1dabc41329607b","abstract_canon_sha256":"6cacfdf61b04f91f85ba394298d22d1edb690af070641444b0399bcbeb8edb81"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:36:08.354509Z","signature_b64":"owWTUaC8r1y5KzS/js1Dsp03Gwf4CSIUB2j1Ea7sbJBb2cfqd102MFLGffvKIkQKhiHuRMKrOFLE3Me2uCdbBg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"4cc445a8895a4f2d9485ace3156f8dd7750db1929cbd9e2ea3088063b8d1f9b0","last_reissued_at":"2026-07-05T11:36:08.354030Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:36:08.354030Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"A Random Matrix Theory Perspective on the Learning Dynamics of Multi-head Latent Attention","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Brandon Reagen, Nandan Kumar Jha","submitted_at":"2025-07-12T20:31:07Z","abstract_excerpt":"In this work, we study how multi-head latent attention (MLA), a popular strategy for compressing key/value memory, affects a transformer's internal capacity during pretraining. Using a lightweight suite of Marchenko-Pastur (MP) diagnostics, we analyze the spectrum of the $W_{Q}W_{K}^\\top$ gram matrix throughout training, comparing three variants: the standard multi-head attention (MHA) baseline, MLA-PreRoPE with rotary applied before compression, and MLA-Decoupled, which shares a single rotary sub-vector across all heads. Our random matrix analysis reveals \\textbf{three key findings:} \\textbf{"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2507.09394","kind":"arxiv","version":1},"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/2507.09394/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":"2507.09394","created_at":"2026-07-05T11:36:08.354087+00:00"},{"alias_kind":"arxiv_version","alias_value":"2507.09394v1","created_at":"2026-07-05T11:36:08.354087+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2507.09394","created_at":"2026-07-05T11:36:08.354087+00:00"},{"alias_kind":"pith_short_12","alias_value":"JTCELKEJLJHS","created_at":"2026-07-05T11:36:08.354087+00:00"},{"alias_kind":"pith_short_16","alias_value":"JTCELKEJLJHS3FEF","created_at":"2026-07-05T11:36:08.354087+00:00"},{"alias_kind":"pith_short_8","alias_value":"JTCELKEJ","created_at":"2026-07-05T11:36:08.354087+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.06826","citing_title":"How Does Attention Help? Insights from Random Matrices on Signal Recovery from Sequence Models","ref_index":9,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/JTCELKEJLJHS3FEFVTRRK34N25","json":"https://pith.science/pith/JTCELKEJLJHS3FEFVTRRK34N25.json","graph_json":"https://pith.science/api/pith-number/JTCELKEJLJHS3FEFVTRRK34N25/graph.json","events_json":"https://pith.science/api/pith-number/JTCELKEJLJHS3FEFVTRRK34N25/events.json","paper":"https://pith.science/paper/JTCELKEJ"},"agent_actions":{"view_html":"https://pith.science/pith/JTCELKEJLJHS3FEFVTRRK34N25","download_json":"https://pith.science/pith/JTCELKEJLJHS3FEFVTRRK34N25.json","view_paper":"https://pith.science/paper/JTCELKEJ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2507.09394&json=true","fetch_graph":"https://pith.science/api/pith-number/JTCELKEJLJHS3FEFVTRRK34N25/graph.json","fetch_events":"https://pith.science/api/pith-number/JTCELKEJLJHS3FEFVTRRK34N25/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/JTCELKEJLJHS3FEFVTRRK34N25/action/timestamp_anchor","attest_storage":"https://pith.science/pith/JTCELKEJLJHS3FEFVTRRK34N25/action/storage_attestation","attest_author":"https://pith.science/pith/JTCELKEJLJHS3FEFVTRRK34N25/action/author_attestation","sign_citation":"https://pith.science/pith/JTCELKEJLJHS3FEFVTRRK34N25/action/citation_signature","submit_replication":"https://pith.science/pith/JTCELKEJLJHS3FEFVTRRK34N25/action/replication_record"}},"created_at":"2026-07-05T11:36:08.354087+00:00","updated_at":"2026-07-05T11:36:08.354087+00:00"}