{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:NPBXIYI36VIAPD4JASPEOLKMQJ","short_pith_number":"pith:NPBXIYI3","schema_version":"1.0","canonical_sha256":"6bc374611bf550078f89049e472d4c82551f0f01bb892246a9dcd7936eac9633","source":{"kind":"arxiv","id":"2602.03685","version":2},"attestation_state":"computed","paper":{"title":"Universal One-third Time Scaling in Learning Peaked Distributions","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","stat.ML"],"primary_cat":"cs.LG","authors_text":"Cengiz Pehlevan, Jeff Gore, Yizhou Liu, Ziming Liu","submitted_at":"2026-02-03T16:06:18Z","abstract_excerpt":"Training large language models (LLMs) is computationally expensive, partly because the loss exhibits slow power-law convergence whose origin remains debatable. Through systematic analysis of toy models and empirical evaluation of LLMs, we show that this behavior can arise intrinsically from the use of softmax and cross-entropy. When learning peaked probability distributions, e.g., next-token distributions, these components generically yield power-law vanishing losses and gradients, regardless of many microscopic details, creating a fundamental optimization bottleneck. This ultimately leads to "},"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":"2602.03685","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-02-03T16:06:18Z","cross_cats_sorted":["cs.AI","stat.ML"],"title_canon_sha256":"e140960ae6811cb551620e3f4b78003a6e0cc024dc03338c82320cebd9e018a8","abstract_canon_sha256":"a6e49dd237934e8a60b769a74a3e1c99acc9743b78cd0d4dcd61263412b1570d"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-02T02:04:13.878549Z","signature_b64":"9t91+TvR6EHCkcpiQQGHZ38ogMb496TDVpRUJGvdJBNwERjWtwp+HOoARqZl+YcYbtF7vqmEGJMNsJAbGclYAA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"6bc374611bf550078f89049e472d4c82551f0f01bb892246a9dcd7936eac9633","last_reissued_at":"2026-06-02T02:04:13.878012Z","signature_status":"signed_v1","first_computed_at":"2026-06-02T02:04:13.878012Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Universal One-third Time Scaling in Learning Peaked Distributions","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","stat.ML"],"primary_cat":"cs.LG","authors_text":"Cengiz Pehlevan, Jeff Gore, Yizhou Liu, Ziming Liu","submitted_at":"2026-02-03T16:06:18Z","abstract_excerpt":"Training large language models (LLMs) is computationally expensive, partly because the loss exhibits slow power-law convergence whose origin remains debatable. Through systematic analysis of toy models and empirical evaluation of LLMs, we show that this behavior can arise intrinsically from the use of softmax and cross-entropy. When learning peaked probability distributions, e.g., next-token distributions, these components generically yield power-law vanishing losses and gradients, regardless of many microscopic details, creating a fundamental optimization bottleneck. This ultimately leads to "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2602.03685","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/2602.03685/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":"2602.03685","created_at":"2026-06-02T02:04:13.878071+00:00"},{"alias_kind":"arxiv_version","alias_value":"2602.03685v2","created_at":"2026-06-02T02:04:13.878071+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2602.03685","created_at":"2026-06-02T02:04:13.878071+00:00"},{"alias_kind":"pith_short_12","alias_value":"NPBXIYI36VIA","created_at":"2026-06-02T02:04:13.878071+00:00"},{"alias_kind":"pith_short_16","alias_value":"NPBXIYI36VIAPD4J","created_at":"2026-06-02T02:04:13.878071+00:00"},{"alias_kind":"pith_short_8","alias_value":"NPBXIYI3","created_at":"2026-06-02T02:04:13.878071+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":3,"internal_anchor_count":3,"sample":[{"citing_arxiv_id":"2605.22341","citing_title":"A Boundary-Layer Mechanism for One-Third Scaling in Online Softmax Classification","ref_index":16,"is_internal_anchor":true},{"citing_arxiv_id":"2605.05686","citing_title":"Attractor Geometry of Transformer Memory: From Conflict Arbitration to Confident Hallucination","ref_index":19,"is_internal_anchor":true},{"citing_arxiv_id":"2605.05686","citing_title":"Attractor Geometry of Transformer Memory: From Conflict Arbitration to Confident Hallucination","ref_index":19,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/NPBXIYI36VIAPD4JASPEOLKMQJ","json":"https://pith.science/pith/NPBXIYI36VIAPD4JASPEOLKMQJ.json","graph_json":"https://pith.science/api/pith-number/NPBXIYI36VIAPD4JASPEOLKMQJ/graph.json","events_json":"https://pith.science/api/pith-number/NPBXIYI36VIAPD4JASPEOLKMQJ/events.json","paper":"https://pith.science/paper/NPBXIYI3"},"agent_actions":{"view_html":"https://pith.science/pith/NPBXIYI36VIAPD4JASPEOLKMQJ","download_json":"https://pith.science/pith/NPBXIYI36VIAPD4JASPEOLKMQJ.json","view_paper":"https://pith.science/paper/NPBXIYI3","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2602.03685&json=true","fetch_graph":"https://pith.science/api/pith-number/NPBXIYI36VIAPD4JASPEOLKMQJ/graph.json","fetch_events":"https://pith.science/api/pith-number/NPBXIYI36VIAPD4JASPEOLKMQJ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/NPBXIYI36VIAPD4JASPEOLKMQJ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/NPBXIYI36VIAPD4JASPEOLKMQJ/action/storage_attestation","attest_author":"https://pith.science/pith/NPBXIYI36VIAPD4JASPEOLKMQJ/action/author_attestation","sign_citation":"https://pith.science/pith/NPBXIYI36VIAPD4JASPEOLKMQJ/action/citation_signature","submit_replication":"https://pith.science/pith/NPBXIYI36VIAPD4JASPEOLKMQJ/action/replication_record"}},"created_at":"2026-06-02T02:04:13.878071+00:00","updated_at":"2026-06-02T02:04:13.878071+00:00"}