{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:OORJVZ5SHIZ4RFXABXVJVDBJKF","short_pith_number":"pith:OORJVZ5S","schema_version":"1.0","canonical_sha256":"73a29ae7b23a33c896e00dea9a8c29517c6f41ff5560a2670a5002e59c4e8a2b","source":{"kind":"arxiv","id":"2406.16793","version":7},"attestation_state":"computed","paper":{"title":"Adam-mini: Use Fewer Learning Rates To Gain More","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Chenwei Wu, Congliang Chen, Diederik P. Kingma, Ruoyu Sun, Tian Ding, Yinyu Ye, Yushun Zhang, Zhi-Quan Luo, Ziniu Li","submitted_at":"2024-06-24T16:56:41Z","abstract_excerpt":"We propose Adam-mini, an optimizer that achieves on par or better performance than AdamW with 50% less memory footprint. Adam-mini reduces memory by cutting down the learning rate resources in Adam (i.e., $1/\\sqrt{v}$). By investigating the Hessian structure of neural nets, we find Adam's $v$ might not function at its full potential as effectively as we expected. We find that $\\geq$ 99.9% of these learning rates in $v$ could be harmlessly removed if we (1) carefully partition the parameters into blocks following our new principle on Hessian structure; (2) assign a single but good learning rate"},"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":"2406.16793","kind":"arxiv","version":7},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-06-24T16:56:41Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"180e4529aa33b034a6053cd8a1ffd3ab8318ad15147a3565453d54090f5b6e0c","abstract_canon_sha256":"8f0700c3fc5ab4964dbcf27de37cbfe64391819584fdc64c115a5fc366a93cf4"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:18:38.789966Z","signature_b64":"5k3lte3mR7hBFO0jdobNvb2OpN0Sajzehe+Ku/92vCo4ps+mtjjj99r+Pn7bpNUsP9A0cERpZRG7A1StaE53Cw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"73a29ae7b23a33c896e00dea9a8c29517c6f41ff5560a2670a5002e59c4e8a2b","last_reissued_at":"2026-07-05T10:18:38.789399Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:18:38.789399Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Adam-mini: Use Fewer Learning Rates To Gain More","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Chenwei Wu, Congliang Chen, Diederik P. Kingma, Ruoyu Sun, Tian Ding, Yinyu Ye, Yushun Zhang, Zhi-Quan Luo, Ziniu Li","submitted_at":"2024-06-24T16:56:41Z","abstract_excerpt":"We propose Adam-mini, an optimizer that achieves on par or better performance than AdamW with 50% less memory footprint. Adam-mini reduces memory by cutting down the learning rate resources in Adam (i.e., $1/\\sqrt{v}$). By investigating the Hessian structure of neural nets, we find Adam's $v$ might not function at its full potential as effectively as we expected. We find that $\\geq$ 99.9% of these learning rates in $v$ could be harmlessly removed if we (1) carefully partition the parameters into blocks following our new principle on Hessian structure; (2) assign a single but good learning rate"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2406.16793","kind":"arxiv","version":7},"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/2406.16793/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":"2406.16793","created_at":"2026-07-05T10:18:38.789459+00:00"},{"alias_kind":"arxiv_version","alias_value":"2406.16793v7","created_at":"2026-07-05T10:18:38.789459+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2406.16793","created_at":"2026-07-05T10:18:38.789459+00:00"},{"alias_kind":"pith_short_12","alias_value":"OORJVZ5SHIZ4","created_at":"2026-07-05T10:18:38.789459+00:00"},{"alias_kind":"pith_short_16","alias_value":"OORJVZ5SHIZ4RFXA","created_at":"2026-07-05T10:18:38.789459+00:00"},{"alias_kind":"pith_short_8","alias_value":"OORJVZ5S","created_at":"2026-07-05T10:18:38.789459+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":18,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2607.05872","citing_title":"No Subspace to Track: Non-Identifiability and Optimizer State in Low-Rank Training","ref_index":15,"is_internal_anchor":true},{"citing_arxiv_id":"2606.08783","citing_title":"OptMuon: Closed-Loop Orthogonalized Momentum Methods for Stochastic Optimization with Zero-Noise Optimality","ref_index":91,"is_internal_anchor":false},{"citing_arxiv_id":"2606.04662","citing_title":"Why Muon Outperforms Adam: A Curvature Perspective","ref_index":198,"is_internal_anchor":false},{"citing_arxiv_id":"2605.22297","citing_title":"One LR Doesn't Fit All: Heavy-Tail Guided Layerwise Learning Rates for LLMs","ref_index":19,"is_internal_anchor":false},{"citing_arxiv_id":"2606.08783","citing_title":"OptMuon: Closed-Loop Orthogonalized Momentum Methods for Stochastic Optimization with Zero-Noise Optimality","ref_index":97,"is_internal_anchor":false},{"citing_arxiv_id":"2605.30960","citing_title":"Revisiting Zeroth-Order Hessian Approximation: A Single-Step Policy Optimization Lens","ref_index":5,"is_internal_anchor":false},{"citing_arxiv_id":"2606.22932","citing_title":"FORGE: Fused On-Register Gradient Elimination for Memory-Efficient LLM Training","ref_index":1,"is_internal_anchor":false},{"citing_arxiv_id":"2603.10067","citing_title":"HTMuon: Improving Muon via Heavy-Tailed Spectral Correction","ref_index":33,"is_internal_anchor":false},{"citing_arxiv_id":"2501.07237","citing_title":"GWT: Scalable Optimizer State Compression for Large Language Model Training","ref_index":53,"is_internal_anchor":false},{"citing_arxiv_id":"2506.16659","citing_title":"Memory-Efficient LLM Pretraining via Minimalist Optimizer Design","ref_index":16,"is_internal_anchor":false},{"citing_arxiv_id":"2605.22297","citing_title":"One LR Doesn't Fit All: Heavy-Tail Guided Layerwise Learning Rates for LLMs","ref_index":20,"is_internal_anchor":false},{"citing_arxiv_id":"2602.00545","citing_title":"Depth, Not Data: An Analysis of Hessian Spectral Bifurcation","ref_index":4,"is_internal_anchor":false},{"citing_arxiv_id":"2605.21177","citing_title":"ChunkFT: Byte-Streamed Optimization for Memory-Efficient Full Fine-Tuning","ref_index":25,"is_internal_anchor":false},{"citing_arxiv_id":"2505.23737","citing_title":"On the Convergence Analysis of Muon","ref_index":27,"is_internal_anchor":false},{"citing_arxiv_id":"2511.16084","citing_title":"SpectralTrain: A Universal Framework for Hyperspectral Image Classification","ref_index":39,"is_internal_anchor":false},{"citing_arxiv_id":"2604.02505","citing_title":"Optimal Projection-Free Adaptive SGD for Matrix Optimization","ref_index":12,"is_internal_anchor":false},{"citing_arxiv_id":"2605.00650","citing_title":"AdaMeZO: Adam-style Zeroth-Order Optimizer for LLM Fine-tuning Without Maintaining the Moments","ref_index":66,"is_internal_anchor":false},{"citing_arxiv_id":"2604.06836","citing_title":"STQuant: Spatio-Temporal Adaptive Framework for Optimizer Quantization in Large Multimodal Model Training","ref_index":34,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/OORJVZ5SHIZ4RFXABXVJVDBJKF","json":"https://pith.science/pith/OORJVZ5SHIZ4RFXABXVJVDBJKF.json","graph_json":"https://pith.science/api/pith-number/OORJVZ5SHIZ4RFXABXVJVDBJKF/graph.json","events_json":"https://pith.science/api/pith-number/OORJVZ5SHIZ4RFXABXVJVDBJKF/events.json","paper":"https://pith.science/paper/OORJVZ5S"},"agent_actions":{"view_html":"https://pith.science/pith/OORJVZ5SHIZ4RFXABXVJVDBJKF","download_json":"https://pith.science/pith/OORJVZ5SHIZ4RFXABXVJVDBJKF.json","view_paper":"https://pith.science/paper/OORJVZ5S","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2406.16793&json=true","fetch_graph":"https://pith.science/api/pith-number/OORJVZ5SHIZ4RFXABXVJVDBJKF/graph.json","fetch_events":"https://pith.science/api/pith-number/OORJVZ5SHIZ4RFXABXVJVDBJKF/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/OORJVZ5SHIZ4RFXABXVJVDBJKF/action/timestamp_anchor","attest_storage":"https://pith.science/pith/OORJVZ5SHIZ4RFXABXVJVDBJKF/action/storage_attestation","attest_author":"https://pith.science/pith/OORJVZ5SHIZ4RFXABXVJVDBJKF/action/author_attestation","sign_citation":"https://pith.science/pith/OORJVZ5SHIZ4RFXABXVJVDBJKF/action/citation_signature","submit_replication":"https://pith.science/pith/OORJVZ5SHIZ4RFXABXVJVDBJKF/action/replication_record"}},"created_at":"2026-07-05T10:18:38.789459+00:00","updated_at":"2026-07-05T10:18:38.789459+00:00"}