{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:COB7HH6HWADJKMQ65BCLONGCVC","short_pith_number":"pith:COB7HH6H","schema_version":"1.0","canonical_sha256":"1383f39fc7b00695321ee844b734c2a8b83ac7bb2d2b5f55e627e1ee8909173d","source":{"kind":"arxiv","id":"2402.19449","version":2},"attestation_state":"computed","paper":{"title":"Heavy-Tailed Class Imbalance and Why Adam Outperforms Gradient Descent on Language Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL","math.OC","stat.ML"],"primary_cat":"cs.LG","authors_text":"Alan Milligan, Alberto Bietti, Frederik Kunstner, Mark Schmidt, Robin Yadav","submitted_at":"2024-02-29T18:47:52Z","abstract_excerpt":"Adam has been shown to outperform gradient descent on large language models by a larger margin than on other tasks, but it is unclear why. We show that a key factor in this performance gap is the heavy-tailed class imbalance found in language tasks. When trained with gradient descent, the loss of infrequent words decreases more slowly than the loss of frequent ones. This leads to a slow decrease on the average loss as most samples come from infrequent words. On the other hand, Adam and sign-based methods are less sensitive to this problem. To establish that this behavior is caused by class imb"},"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":"2402.19449","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2024-02-29T18:47:52Z","cross_cats_sorted":["cs.CL","math.OC","stat.ML"],"title_canon_sha256":"fbf939afc9c4ddcbb6611a0f3a3b30ce6a28d766294be4b33e869daa812d0713","abstract_canon_sha256":"49ca9f6ef3a45a0a529a374ced6bdee2057accf1be1a040704d6a0022e66f62a"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:43:08.867486Z","signature_b64":"ExjkzUVheMDyJS9ATR7Kpe0VhYmEQo3kIAhzwvM0yLhJ2cFYWvKhnwiQV1o1yCSBAcFHFW355sq50Oo0BigYBA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"1383f39fc7b00695321ee844b734c2a8b83ac7bb2d2b5f55e627e1ee8909173d","last_reissued_at":"2026-07-05T08:43:08.867070Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:43:08.867070Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Heavy-Tailed Class Imbalance and Why Adam Outperforms Gradient Descent on Language Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL","math.OC","stat.ML"],"primary_cat":"cs.LG","authors_text":"Alan Milligan, Alberto Bietti, Frederik Kunstner, Mark Schmidt, Robin Yadav","submitted_at":"2024-02-29T18:47:52Z","abstract_excerpt":"Adam has been shown to outperform gradient descent on large language models by a larger margin than on other tasks, but it is unclear why. We show that a key factor in this performance gap is the heavy-tailed class imbalance found in language tasks. When trained with gradient descent, the loss of infrequent words decreases more slowly than the loss of frequent ones. This leads to a slow decrease on the average loss as most samples come from infrequent words. On the other hand, Adam and sign-based methods are less sensitive to this problem. To establish that this behavior is caused by class imb"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2402.19449","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/2402.19449/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":"2402.19449","created_at":"2026-07-05T08:43:08.867125+00:00"},{"alias_kind":"arxiv_version","alias_value":"2402.19449v2","created_at":"2026-07-05T08:43:08.867125+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2402.19449","created_at":"2026-07-05T08:43:08.867125+00:00"},{"alias_kind":"pith_short_12","alias_value":"COB7HH6HWADJ","created_at":"2026-07-05T08:43:08.867125+00:00"},{"alias_kind":"pith_short_16","alias_value":"COB7HH6HWADJKMQ6","created_at":"2026-07-05T08:43:08.867125+00:00"},{"alias_kind":"pith_short_8","alias_value":"COB7HH6H","created_at":"2026-07-05T08:43:08.867125+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":3,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.09253","citing_title":"Cornerstones or Stumbling Blocks? Deciphering the Rock Tokens in On-Policy Distillation","ref_index":31,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09253","citing_title":"Cornerstones or Stumbling Blocks? Deciphering the Rock Tokens in On-Policy Distillation","ref_index":31,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09253","citing_title":"Cornerstones or Stumbling Blocks? Deciphering the Rock Tokens in On-Policy Distillation","ref_index":16,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/COB7HH6HWADJKMQ65BCLONGCVC","json":"https://pith.science/pith/COB7HH6HWADJKMQ65BCLONGCVC.json","graph_json":"https://pith.science/api/pith-number/COB7HH6HWADJKMQ65BCLONGCVC/graph.json","events_json":"https://pith.science/api/pith-number/COB7HH6HWADJKMQ65BCLONGCVC/events.json","paper":"https://pith.science/paper/COB7HH6H"},"agent_actions":{"view_html":"https://pith.science/pith/COB7HH6HWADJKMQ65BCLONGCVC","download_json":"https://pith.science/pith/COB7HH6HWADJKMQ65BCLONGCVC.json","view_paper":"https://pith.science/paper/COB7HH6H","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2402.19449&json=true","fetch_graph":"https://pith.science/api/pith-number/COB7HH6HWADJKMQ65BCLONGCVC/graph.json","fetch_events":"https://pith.science/api/pith-number/COB7HH6HWADJKMQ65BCLONGCVC/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/COB7HH6HWADJKMQ65BCLONGCVC/action/timestamp_anchor","attest_storage":"https://pith.science/pith/COB7HH6HWADJKMQ65BCLONGCVC/action/storage_attestation","attest_author":"https://pith.science/pith/COB7HH6HWADJKMQ65BCLONGCVC/action/author_attestation","sign_citation":"https://pith.science/pith/COB7HH6HWADJKMQ65BCLONGCVC/action/citation_signature","submit_replication":"https://pith.science/pith/COB7HH6HWADJKMQ65BCLONGCVC/action/replication_record"}},"created_at":"2026-07-05T08:43:08.867125+00:00","updated_at":"2026-07-05T08:43:08.867125+00:00"}