{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:NKSNQO5YLYO2EIZHKUQRUN773X","short_pith_number":"pith:NKSNQO5Y","schema_version":"1.0","canonical_sha256":"6aa4d83bb85e1da2232755211a37ffddd2b97da6dda1e4f470fe0d46bb55feed","source":{"kind":"arxiv","id":"2506.05447","version":2},"attestation_state":"computed","paper":{"title":"Training Dynamics Underlying Language Model Scaling Laws: Loss Deceleration and Zero-Sum Learning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Andrei Mircea, Ekaterina Lobacheva, Irina Rish, Milind Naphade, Nima Chitsazan, Sambit Sahu, Supriyo Chakraborty","submitted_at":"2025-06-05T15:18:35Z","abstract_excerpt":"This work aims to understand how scaling improves language models, specifically in terms of training dynamics. We find that language models undergo loss deceleration early in training; an abrupt slowdown in the rate of loss improvement, resulting in piecewise linear behaviour of the loss curve in log-log space. Scaling up the model mitigates this transition by (1) decreasing the loss at which deceleration occurs, and (2) improving the log-log rate of loss improvement after deceleration. We attribute loss deceleration to a type of degenerate training dynamics we term zero-sum learning (ZSL). In"},"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":"2506.05447","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2025-06-05T15:18:35Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"dca44d6c5ec302965d45eb1f7e214b72394537877dafc9efd5cafcd2c43e2592","abstract_canon_sha256":"e1496d502b117559b4ffa1d5b694ce269d1fe0193fdf61d9ffeb9276101296eb"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:37:22.835780Z","signature_b64":"bHBcmz/tZK8oKX9qtfQTFLoF8rkTlXaLUaNx5GR+KTPql6DWU5NajJb/ne2c0sQeC7kdiNKWv0jCIW91QJg/BQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"6aa4d83bb85e1da2232755211a37ffddd2b97da6dda1e4f470fe0d46bb55feed","last_reissued_at":"2026-07-05T11:37:22.835270Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:37:22.835270Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Training Dynamics Underlying Language Model Scaling Laws: Loss Deceleration and Zero-Sum Learning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Andrei Mircea, Ekaterina Lobacheva, Irina Rish, Milind Naphade, Nima Chitsazan, Sambit Sahu, Supriyo Chakraborty","submitted_at":"2025-06-05T15:18:35Z","abstract_excerpt":"This work aims to understand how scaling improves language models, specifically in terms of training dynamics. We find that language models undergo loss deceleration early in training; an abrupt slowdown in the rate of loss improvement, resulting in piecewise linear behaviour of the loss curve in log-log space. Scaling up the model mitigates this transition by (1) decreasing the loss at which deceleration occurs, and (2) improving the log-log rate of loss improvement after deceleration. We attribute loss deceleration to a type of degenerate training dynamics we term zero-sum learning (ZSL). In"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2506.05447","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/2506.05447/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":"2506.05447","created_at":"2026-07-05T11:37:22.835334+00:00"},{"alias_kind":"arxiv_version","alias_value":"2506.05447v2","created_at":"2026-07-05T11:37:22.835334+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2506.05447","created_at":"2026-07-05T11:37:22.835334+00:00"},{"alias_kind":"pith_short_12","alias_value":"NKSNQO5YLYO2","created_at":"2026-07-05T11:37:22.835334+00:00"},{"alias_kind":"pith_short_16","alias_value":"NKSNQO5YLYO2EIZH","created_at":"2026-07-05T11:37:22.835334+00:00"},{"alias_kind":"pith_short_8","alias_value":"NKSNQO5Y","created_at":"2026-07-05T11:37:22.835334+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2607.25271","citing_title":"Bridging Compute- and Data-Optimal Pretraining","ref_index":80,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/NKSNQO5YLYO2EIZHKUQRUN773X","json":"https://pith.science/pith/NKSNQO5YLYO2EIZHKUQRUN773X.json","graph_json":"https://pith.science/api/pith-number/NKSNQO5YLYO2EIZHKUQRUN773X/graph.json","events_json":"https://pith.science/api/pith-number/NKSNQO5YLYO2EIZHKUQRUN773X/events.json","paper":"https://pith.science/paper/NKSNQO5Y"},"agent_actions":{"view_html":"https://pith.science/pith/NKSNQO5YLYO2EIZHKUQRUN773X","download_json":"https://pith.science/pith/NKSNQO5YLYO2EIZHKUQRUN773X.json","view_paper":"https://pith.science/paper/NKSNQO5Y","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2506.05447&json=true","fetch_graph":"https://pith.science/api/pith-number/NKSNQO5YLYO2EIZHKUQRUN773X/graph.json","fetch_events":"https://pith.science/api/pith-number/NKSNQO5YLYO2EIZHKUQRUN773X/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/NKSNQO5YLYO2EIZHKUQRUN773X/action/timestamp_anchor","attest_storage":"https://pith.science/pith/NKSNQO5YLYO2EIZHKUQRUN773X/action/storage_attestation","attest_author":"https://pith.science/pith/NKSNQO5YLYO2EIZHKUQRUN773X/action/author_attestation","sign_citation":"https://pith.science/pith/NKSNQO5YLYO2EIZHKUQRUN773X/action/citation_signature","submit_replication":"https://pith.science/pith/NKSNQO5YLYO2EIZHKUQRUN773X/action/replication_record"}},"created_at":"2026-07-05T11:37:22.835334+00:00","updated_at":"2026-07-05T11:37:22.835334+00:00"}