{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:5GLYRQWISBMTTF3GN5S5FKBQ66","short_pith_number":"pith:5GLYRQWI","schema_version":"1.0","canonical_sha256":"e99788c2c890593997666f65d2a830f79dc294613e3e1f0b9a257e5a4c8ad390","source":{"kind":"arxiv","id":"2506.16659","version":3},"attestation_state":"computed","paper":{"title":"Memory-Efficient LLM Pretraining via Minimalist Optimizer Design","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","math.OC"],"primary_cat":"cs.LG","authors_text":"Andi Han, Athanasios Glentis, Jiaxiang Li, Mingyi Hong","submitted_at":"2025-06-20T00:10:35Z","abstract_excerpt":"Training large language models (LLMs) relies on adaptive optimizers such as Adam, which introduce extra operations and require significantly more memory to maintain first- and second-order moments than SGD. While recent works such as GaLore, Fira and APOLLO have proposed state-compressed memory-efficient variants, a fundamental question remains: What are the minimum modifications to plain SGD needed to match state-of-the-art pretraining performance? We systematically investigate this question using a bottom-up approach, and identify two simple yet highly (memory- and compute-) efficient techni"},"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":"2506.16659","kind":"arxiv","version":3},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2025-06-20T00:10:35Z","cross_cats_sorted":["cs.AI","math.OC"],"title_canon_sha256":"e23700cd447f3fb289457403c1de9e4e4ccd2b6cd75f0238eda93186941fe2ea","abstract_canon_sha256":"fb145049edde77775fb72af8b6b40a4ae2381029475cb35bc5fca05534044295"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-22T01:03:43.627522Z","signature_b64":"TnpRBpAGahUvU9uiv00WD6hKQmwHKvxcH4WolPu/mcTPfCN0IkEqBYw2MY2+OKcHk1mB7EEZHOgIJHhW0ianAA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"e99788c2c890593997666f65d2a830f79dc294613e3e1f0b9a257e5a4c8ad390","last_reissued_at":"2026-05-22T01:03:43.626630Z","signature_status":"signed_v1","first_computed_at":"2026-05-22T01:03:43.626630Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Memory-Efficient LLM Pretraining via Minimalist Optimizer Design","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","math.OC"],"primary_cat":"cs.LG","authors_text":"Andi Han, Athanasios Glentis, Jiaxiang Li, Mingyi Hong","submitted_at":"2025-06-20T00:10:35Z","abstract_excerpt":"Training large language models (LLMs) relies on adaptive optimizers such as Adam, which introduce extra operations and require significantly more memory to maintain first- and second-order moments than SGD. While recent works such as GaLore, Fira and APOLLO have proposed state-compressed memory-efficient variants, a fundamental question remains: What are the minimum modifications to plain SGD needed to match state-of-the-art pretraining performance? We systematically investigate this question using a bottom-up approach, and identify two simple yet highly (memory- and compute-) efficient techni"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2506.16659","kind":"arxiv","version":3},"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/2506.16659/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":"2506.16659","created_at":"2026-05-22T01:03:43.626757+00:00"},{"alias_kind":"arxiv_version","alias_value":"2506.16659v3","created_at":"2026-05-22T01:03:43.626757+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2506.16659","created_at":"2026-05-22T01:03:43.626757+00:00"},{"alias_kind":"pith_short_12","alias_value":"5GLYRQWISBMT","created_at":"2026-05-22T01:03:43.626757+00:00"},{"alias_kind":"pith_short_16","alias_value":"5GLYRQWISBMTTF3G","created_at":"2026-05-22T01:03:43.626757+00:00"},{"alias_kind":"pith_short_8","alias_value":"5GLYRQWI","created_at":"2026-05-22T01:03:43.626757+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":5,"internal_anchor_count":5,"sample":[{"citing_arxiv_id":"2605.18106","citing_title":"Symmetry-Compatible Principle for Optimizer Design: Embeddings, LM Heads, SwiGLU MLPs, and MoE Routers","ref_index":51,"is_internal_anchor":true},{"citing_arxiv_id":"2509.11983","citing_title":"Low-rank Orthogonalization for Large-scale Matrix Optimization with Applications to Foundation Model Training","ref_index":20,"is_internal_anchor":true},{"citing_arxiv_id":"2603.28254","citing_title":"MuonEq: Balancing Before Orthogonalization with Lightweight Equilibration","ref_index":40,"is_internal_anchor":true},{"citing_arxiv_id":"2605.04418","citing_title":"Demystifying Manifold Constraints in LLM Pre-training","ref_index":42,"is_internal_anchor":true},{"citing_arxiv_id":"2605.04711","citing_title":"Budget-aware Auto Optimizer Configurator","ref_index":13,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/5GLYRQWISBMTTF3GN5S5FKBQ66","json":"https://pith.science/pith/5GLYRQWISBMTTF3GN5S5FKBQ66.json","graph_json":"https://pith.science/api/pith-number/5GLYRQWISBMTTF3GN5S5FKBQ66/graph.json","events_json":"https://pith.science/api/pith-number/5GLYRQWISBMTTF3GN5S5FKBQ66/events.json","paper":"https://pith.science/paper/5GLYRQWI"},"agent_actions":{"view_html":"https://pith.science/pith/5GLYRQWISBMTTF3GN5S5FKBQ66","download_json":"https://pith.science/pith/5GLYRQWISBMTTF3GN5S5FKBQ66.json","view_paper":"https://pith.science/paper/5GLYRQWI","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2506.16659&json=true","fetch_graph":"https://pith.science/api/pith-number/5GLYRQWISBMTTF3GN5S5FKBQ66/graph.json","fetch_events":"https://pith.science/api/pith-number/5GLYRQWISBMTTF3GN5S5FKBQ66/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/5GLYRQWISBMTTF3GN5S5FKBQ66/action/timestamp_anchor","attest_storage":"https://pith.science/pith/5GLYRQWISBMTTF3GN5S5FKBQ66/action/storage_attestation","attest_author":"https://pith.science/pith/5GLYRQWISBMTTF3GN5S5FKBQ66/action/author_attestation","sign_citation":"https://pith.science/pith/5GLYRQWISBMTTF3GN5S5FKBQ66/action/citation_signature","submit_replication":"https://pith.science/pith/5GLYRQWISBMTTF3GN5S5FKBQ66/action/replication_record"}},"created_at":"2026-05-22T01:03:43.626757+00:00","updated_at":"2026-05-22T01:03:43.626757+00:00"}