{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2023:SO76PEDCYF7C2AJR32CVFSZNOP","short_pith_number":"pith:SO76PEDC","schema_version":"1.0","canonical_sha256":"93bfe79062c17e2d0131de8552cb2d73d50766b387be15adbc66e345222c17e5","source":{"kind":"arxiv","id":"2304.09871","version":2},"attestation_state":"computed","paper":{"title":"A Theory on Adam Instability in Large-Scale Machine Learning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","math.OC"],"primary_cat":"cs.LG","authors_text":"Andrew Poulton, Binh Tang, David Esiobu, Diana Liskovich, Igor Molybog, Melanie Kambadur, Moya Chen, Naman Goyal, Peter Albert, Punit Singh Koura, Puxin Xu, Ruan Silva, Sharan Narang, Stephen Roller, Susan Zhang, Yuchen Zhang, Zachary DeVito","submitted_at":"2023-04-19T06:15:11Z","abstract_excerpt":"We present a theory for the previously unexplained divergent behavior noticed in the training of large language models. We argue that the phenomenon is an artifact of the dominant optimization algorithm used for training, called Adam. We observe that Adam can enter a state in which the parameter update vector has a relatively large norm and is essentially uncorrelated with the direction of descent on the training loss landscape, leading to divergence. This artifact is more likely to be observed in the training of a deep model with a large batch size, which is the typical setting of large-scale"},"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":"2304.09871","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2023-04-19T06:15:11Z","cross_cats_sorted":["cs.AI","math.OC"],"title_canon_sha256":"6553f380d521f4433f5c6b03ae6394880b96859825918bcbb65930d100ff853c","abstract_canon_sha256":"a21c0b5d276017e7c3570540e757eeedc76ce8b44705c1c0294914c1550602ac"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T06:04:14.377568Z","signature_b64":"d5D9FPu4l9KIfs7uNfJVVsa/6cgniRvcBkWYI2sS+4TpPHzst0G3vIgXiZ/7wh7nm2n8EY20b7/ycz0cxxe0BQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"93bfe79062c17e2d0131de8552cb2d73d50766b387be15adbc66e345222c17e5","last_reissued_at":"2026-07-05T06:04:14.377109Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T06:04:14.377109Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"A Theory on Adam Instability in Large-Scale Machine Learning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","math.OC"],"primary_cat":"cs.LG","authors_text":"Andrew Poulton, Binh Tang, David Esiobu, Diana Liskovich, Igor Molybog, Melanie Kambadur, Moya Chen, Naman Goyal, Peter Albert, Punit Singh Koura, Puxin Xu, Ruan Silva, Sharan Narang, Stephen Roller, Susan Zhang, Yuchen Zhang, Zachary DeVito","submitted_at":"2023-04-19T06:15:11Z","abstract_excerpt":"We present a theory for the previously unexplained divergent behavior noticed in the training of large language models. We argue that the phenomenon is an artifact of the dominant optimization algorithm used for training, called Adam. We observe that Adam can enter a state in which the parameter update vector has a relatively large norm and is essentially uncorrelated with the direction of descent on the training loss landscape, leading to divergence. This artifact is more likely to be observed in the training of a deep model with a large batch size, which is the typical setting of large-scale"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2304.09871","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/2304.09871/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":"2304.09871","created_at":"2026-07-05T06:04:14.377171+00:00"},{"alias_kind":"arxiv_version","alias_value":"2304.09871v2","created_at":"2026-07-05T06:04:14.377171+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2304.09871","created_at":"2026-07-05T06:04:14.377171+00:00"},{"alias_kind":"pith_short_12","alias_value":"SO76PEDCYF7C","created_at":"2026-07-05T06:04:14.377171+00:00"},{"alias_kind":"pith_short_16","alias_value":"SO76PEDCYF7C2AJR","created_at":"2026-07-05T06:04:14.377171+00:00"},{"alias_kind":"pith_short_8","alias_value":"SO76PEDC","created_at":"2026-07-05T06:04:14.377171+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":12,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2607.01678","citing_title":"SCAPE: Accurate and Efficient LLM Training with Extreme Sparse Communication","ref_index":27,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06152","citing_title":"Grokking or Glitching? How Low-Precision Drives Slingshot Loss Spikes","ref_index":32,"is_internal_anchor":false},{"citing_arxiv_id":"2605.25704","citing_title":"PowLU: An Activation Function for Stable Pre-Training of LLMs","ref_index":15,"is_internal_anchor":false},{"citing_arxiv_id":"2501.07237","citing_title":"GWT: Scalable Optimizer State Compression for Large Language Model Training","ref_index":30,"is_internal_anchor":false},{"citing_arxiv_id":"2507.06261","citing_title":"Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities","ref_index":55,"is_internal_anchor":false},{"citing_arxiv_id":"2510.04212","citing_title":"Why Low-Precision Transformer Training Fails: An Analysis on Flash Attention","ref_index":17,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06152","citing_title":"Grokking or Glitching? How Low-Precision Drives Slingshot Loss Spikes","ref_index":24,"is_internal_anchor":false},{"citing_arxiv_id":"2506.13585","citing_title":"MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention","ref_index":23,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09176","citing_title":"Navigating LLM Valley: From AdamW to Memory-Efficient and Matrix-Based Optimizers","ref_index":25,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06152","citing_title":"Grokking or Glitching? How Low-Precision Drives Slingshot Loss Spikes","ref_index":24,"is_internal_anchor":false},{"citing_arxiv_id":"2412.20404","citing_title":"Open-Sora: Democratizing Efficient Video Production for All","ref_index":21,"is_internal_anchor":false},{"citing_arxiv_id":"2505.14683","citing_title":"Emerging Properties in Unified Multimodal Pretraining","ref_index":52,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/SO76PEDCYF7C2AJR32CVFSZNOP","json":"https://pith.science/pith/SO76PEDCYF7C2AJR32CVFSZNOP.json","graph_json":"https://pith.science/api/pith-number/SO76PEDCYF7C2AJR32CVFSZNOP/graph.json","events_json":"https://pith.science/api/pith-number/SO76PEDCYF7C2AJR32CVFSZNOP/events.json","paper":"https://pith.science/paper/SO76PEDC"},"agent_actions":{"view_html":"https://pith.science/pith/SO76PEDCYF7C2AJR32CVFSZNOP","download_json":"https://pith.science/pith/SO76PEDCYF7C2AJR32CVFSZNOP.json","view_paper":"https://pith.science/paper/SO76PEDC","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2304.09871&json=true","fetch_graph":"https://pith.science/api/pith-number/SO76PEDCYF7C2AJR32CVFSZNOP/graph.json","fetch_events":"https://pith.science/api/pith-number/SO76PEDCYF7C2AJR32CVFSZNOP/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/SO76PEDCYF7C2AJR32CVFSZNOP/action/timestamp_anchor","attest_storage":"https://pith.science/pith/SO76PEDCYF7C2AJR32CVFSZNOP/action/storage_attestation","attest_author":"https://pith.science/pith/SO76PEDCYF7C2AJR32CVFSZNOP/action/author_attestation","sign_citation":"https://pith.science/pith/SO76PEDCYF7C2AJR32CVFSZNOP/action/citation_signature","submit_replication":"https://pith.science/pith/SO76PEDCYF7C2AJR32CVFSZNOP/action/replication_record"}},"created_at":"2026-07-05T06:04:14.377171+00:00","updated_at":"2026-07-05T06:04:14.377171+00:00"}