{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:N7BATLN5YJWMBDF5XOT3BUFQ7X","short_pith_number":"pith:N7BATLN5","schema_version":"1.0","canonical_sha256":"6fc209adbdc26cc08cbdbba7b0d0b0fdd7df22204eaa244b9c6c097d48c04f69","source":{"kind":"arxiv","id":"2503.21676","version":2},"attestation_state":"computed","paper":{"title":"How do language models learn facts? Dynamics, curricula and hallucinations","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Andrew Lampinen, J\\\"org Bornschein, Nicolas Zucchet, Razvan Pascanu, Soham De, Stephanie Chan","submitted_at":"2025-03-27T16:43:45Z","abstract_excerpt":"Large language models accumulate vast knowledge during pre-training, yet the dynamics governing this acquisition remain poorly understood. This work investigates the learning dynamics of language models on a synthetic factual recall task, uncovering three key findings: First, language models learn in three phases, exhibiting a performance plateau before acquiring precise factual knowledge. Mechanistically, this plateau coincides with the formation of attention-based circuits that support recall. Second, the training data distribution significantly impacts learning dynamics, as imbalanced distr"},"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":"2503.21676","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2025-03-27T16:43:45Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"c4d6664ada388834dad9f85b90c76255a8a5de9191822c9cd57584dfb67f459d","abstract_canon_sha256":"9d313aff0067a1316b97e7a08ae2c8149650759e2ba645d0d1df981f9af8f5ff"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:42:28.026730Z","signature_b64":"G+G/02kFO7NisWu761E7kr3d4At6QPbuXryrUuu3GSMVScrK1ZW0hYrnBMQpTlmho6B8jEaQVAjCrN+6jhWbAA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"6fc209adbdc26cc08cbdbba7b0d0b0fdd7df22204eaa244b9c6c097d48c04f69","last_reissued_at":"2026-07-05T11:42:28.026183Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:42:28.026183Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"How do language models learn facts? Dynamics, curricula and hallucinations","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Andrew Lampinen, J\\\"org Bornschein, Nicolas Zucchet, Razvan Pascanu, Soham De, Stephanie Chan","submitted_at":"2025-03-27T16:43:45Z","abstract_excerpt":"Large language models accumulate vast knowledge during pre-training, yet the dynamics governing this acquisition remain poorly understood. This work investigates the learning dynamics of language models on a synthetic factual recall task, uncovering three key findings: First, language models learn in three phases, exhibiting a performance plateau before acquiring precise factual knowledge. Mechanistically, this plateau coincides with the formation of attention-based circuits that support recall. Second, the training data distribution significantly impacts learning dynamics, as imbalanced distr"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2503.21676","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/2503.21676/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":"2503.21676","created_at":"2026-07-05T11:42:28.026242+00:00"},{"alias_kind":"arxiv_version","alias_value":"2503.21676v2","created_at":"2026-07-05T11:42:28.026242+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2503.21676","created_at":"2026-07-05T11:42:28.026242+00:00"},{"alias_kind":"pith_short_12","alias_value":"N7BATLN5YJWM","created_at":"2026-07-05T11:42:28.026242+00:00"},{"alias_kind":"pith_short_16","alias_value":"N7BATLN5YJWMBDF5","created_at":"2026-07-05T11:42:28.026242+00:00"},{"alias_kind":"pith_short_8","alias_value":"N7BATLN5","created_at":"2026-07-05T11:42:28.026242+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":9,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.26050","citing_title":"Natural Ungrokking: Asymmetric Control of Which Rules Survive Pretraining","ref_index":48,"is_internal_anchor":false},{"citing_arxiv_id":"2606.26749","citing_title":"Structure Before Collapse: Transient semantic geometry in next-token prediction","ref_index":14,"is_internal_anchor":false},{"citing_arxiv_id":"2605.27564","citing_title":"The Future of Facts: Tracing the Factual Generation-Verification Gap","ref_index":33,"is_internal_anchor":false},{"citing_arxiv_id":"2510.26745","citing_title":"Deep sequence models tend to memorize geometrically; it is unclear why","ref_index":211,"is_internal_anchor":false},{"citing_arxiv_id":"2605.20005","citing_title":"Fine-Tuning Without Forgetting via Loss-Adaptive Learning Rates","ref_index":76,"is_internal_anchor":false},{"citing_arxiv_id":"2509.13316","citing_title":"Do Activation Verbalization Methods Convey Privileged Information?","ref_index":57,"is_internal_anchor":false},{"citing_arxiv_id":"2510.02370","citing_title":"How Training Data Shapes the Use of Parametric and In-Context Knowledge in Language Models","ref_index":20,"is_internal_anchor":false},{"citing_arxiv_id":"2604.08519","citing_title":"Cram Less to Fit More: Training Data Pruning Improves Memorization of Facts","ref_index":103,"is_internal_anchor":false},{"citing_arxiv_id":"2604.15574","citing_title":"Why Fine-Tuning Encourages Hallucinations and How to Fix It","ref_index":4,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/N7BATLN5YJWMBDF5XOT3BUFQ7X","json":"https://pith.science/pith/N7BATLN5YJWMBDF5XOT3BUFQ7X.json","graph_json":"https://pith.science/api/pith-number/N7BATLN5YJWMBDF5XOT3BUFQ7X/graph.json","events_json":"https://pith.science/api/pith-number/N7BATLN5YJWMBDF5XOT3BUFQ7X/events.json","paper":"https://pith.science/paper/N7BATLN5"},"agent_actions":{"view_html":"https://pith.science/pith/N7BATLN5YJWMBDF5XOT3BUFQ7X","download_json":"https://pith.science/pith/N7BATLN5YJWMBDF5XOT3BUFQ7X.json","view_paper":"https://pith.science/paper/N7BATLN5","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2503.21676&json=true","fetch_graph":"https://pith.science/api/pith-number/N7BATLN5YJWMBDF5XOT3BUFQ7X/graph.json","fetch_events":"https://pith.science/api/pith-number/N7BATLN5YJWMBDF5XOT3BUFQ7X/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/N7BATLN5YJWMBDF5XOT3BUFQ7X/action/timestamp_anchor","attest_storage":"https://pith.science/pith/N7BATLN5YJWMBDF5XOT3BUFQ7X/action/storage_attestation","attest_author":"https://pith.science/pith/N7BATLN5YJWMBDF5XOT3BUFQ7X/action/author_attestation","sign_citation":"https://pith.science/pith/N7BATLN5YJWMBDF5XOT3BUFQ7X/action/citation_signature","submit_replication":"https://pith.science/pith/N7BATLN5YJWMBDF5XOT3BUFQ7X/action/replication_record"}},"created_at":"2026-07-05T11:42:28.026242+00:00","updated_at":"2026-07-05T11:42:28.026242+00:00"}