{"paper":{"title":"GQA-{\\mu}P: The maximal parameterization update for grouped query attention","license":"http://creativecommons.org/licenses/by/4.0/","headline":"A modified spectral norm for non-full-rank matrices lets maximal update parameterization apply to grouped-query attention.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Alexander Moreno, Daria Soboleva, Eric Xing, Huijuan Wang, Joel Hestness, Kyle R. Chickering, Mengxi Wu, Muhao Chen, Xuezhe Ma, Zhengzhong Liu","submitted_at":"2026-05-14T18:03:16Z","abstract_excerpt":"Hyperparameter transfer across model architectures dramatically reduces the amount of compute necessary for tuning large language models (LLMs). The maximal update parameterization ({\\mu}P) ensures transfer through principled mathematical analysis but can be challenging to derive for new model architectures. Building on the spectral feature-learning view of Yang et al. (2023a), we make two advances. First, we promote spectral norm conditions on the weights from a heuristic to the definition of feature learning, and as a consequence arrive at the Complete-P depth and weight-decay scalings witho"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We demonstrate the efficacy of our theoretical derivations by showing learning rate transfer across the GQA repetition hyperparameter as well as experiments regarding transfer over weight decay.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The modified spectral norm preserves the valid scaling law of network weights when weight matrices are not full rank; this premise is invoked to enable the GQA derivation and is stated as the key technical step after promoting spectral conditions to a definition.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Derives μP scalings for GQA via promoted spectral-norm definition of feature learning and a modified norm preserving scaling laws for non-full-rank matrices, with experiments showing learning-rate transfer.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A modified spectral norm for non-full-rank matrices lets maximal update parameterization apply to grouped-query attention.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"9b97e1d39f8720e1de5e39e19a35524effc84db90983169c75a8dc14a689d81f"},"source":{"id":"2605.15290","kind":"arxiv","version":1},"verdict":{"id":"8284b6d5-b2d5-44f0-a2fd-d986c81ac348","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T16:31:30.644410Z","strongest_claim":"We demonstrate the efficacy of our theoretical derivations by showing learning rate transfer across the GQA repetition hyperparameter as well as experiments regarding transfer over weight decay.","one_line_summary":"Derives μP scalings for GQA via promoted spectral-norm definition of feature learning and a modified norm preserving scaling laws for non-full-rank matrices, with experiments showing learning-rate transfer.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The modified spectral norm preserves the valid scaling law of network weights when weight matrices are not full rank; this premise is invoked to enable the GQA derivation and is stated as the key technical step after promoting spectral conditions to a definition.","pith_extraction_headline":"A modified spectral norm for non-full-rank matrices lets maximal update parameterization apply to grouped-query attention."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.15290/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_title_agreement","ran_at":"2026-05-19T17:01:18.316486Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T16:45:40.492995Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"claim_evidence","ran_at":"2026-05-19T14:41:54.239751Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T13:33:22.786996Z","status":"skipped","version":"1.0.0","findings_count":0}],"snapshot_sha256":"96c1b083a4d1493f805e127ba21ec7c2e55e2f77d7ff72c04c27d138b4c58472"},"references":{"count":30,"sample":[{"doi":"","year":null,"title":"GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints","work_id":"b73ad5b2-e553-4c71-b0c9-67e67ba7b158","ref_index":1,"cited_arxiv_id":"2305.13245","is_internal_anchor":true},{"doi":"","year":null,"title":"Why do we need weight decay in modern deep learning? ArXiv, abs/2310.04415","work_id":"c97a49ec-f435-4fcd-812a-ac270ff17004","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Power lines: Scaling laws for weight decay and batch size in llm pre-training","work_id":"e2555467-0d85-4bbb-8e73-ad73ad65dc98","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Cerebras-gpt: Open compute- optimal language models trained on the cerebras wafer- scale cluster","work_id":"146665a8-e124-465b-aa7f-84693124b019","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Don’t be lazy: CompleteP enables compute- efficient deep transformers, January 2026","work_id":"85f11780-ed20-4881-8d31-bb0834b58027","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":30,"snapshot_sha256":"21846ac245695c053d53eb777e3260b88e0070308a0f70598da4853a8be9574e","internal_anchors":8},"formal_canon":{"evidence_count":1,"snapshot_sha256":"a3c123c89a3bcfc1e6766430c4cd2dc48121a8cd654203f36a38c2dc2bd07741"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}