{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:KLPQSYFTTAOMX3NQD4P5HIUFOY","short_pith_number":"pith:KLPQSYFT","schema_version":"1.0","canonical_sha256":"52df0960b3981ccbedb01f1fd3a2857626797fbf4c887b29a5ff3145ae1ddfe7","source":{"kind":"arxiv","id":"2603.12901","version":1},"attestation_state":"computed","paper":{"title":"A theory of learning data statistics in diffusion models, from easy to hard","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cond-mat.dis-nn","cs.IT","cs.LG","math.IT"],"primary_cat":"stat.ML","authors_text":"Claudia Merger, Lorenzo Bardone, Sebastian Goldt","submitted_at":"2026-03-13T11:07:01Z","abstract_excerpt":"While diffusion models have emerged as a powerful class of generative models, their learning dynamics remain poorly understood. We address this issue first by empirically showing that standard diffusion models trained on natural images exhibit a distributional simplicity bias, learning simple, pair-wise input statistics before specializing to higher-order correlations. We reproduce this behaviour in simple denoisers trained on a minimal data model, the mixed cumulant model, where we precisely control both pair-wise and higher-order correlations of the inputs. We identify a scalar invariant of "},"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":"2603.12901","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"stat.ML","submitted_at":"2026-03-13T11:07:01Z","cross_cats_sorted":["cond-mat.dis-nn","cs.IT","cs.LG","math.IT"],"title_canon_sha256":"1e96804b1adcd5afc71a4a382bc4289c9e1040d6dd44a992d17a0615d01d3940","abstract_canon_sha256":"8148d4f2d20641b0afa4526c6f507da2bd55cbfa3df61af5b9e54e520885feba"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-21T02:05:00.972041Z","signature_b64":"Ckg7UwvY/KrKbEi5uWwf4Xn6zBNxpdnIf44GIswPf/o+NV7wN8YgPsy9vYPEBCkWjz7MUfzrlAOBjSg8eUWtAw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"52df0960b3981ccbedb01f1fd3a2857626797fbf4c887b29a5ff3145ae1ddfe7","last_reissued_at":"2026-05-21T02:05:00.971207Z","signature_status":"signed_v1","first_computed_at":"2026-05-21T02:05:00.971207Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"A theory of learning data statistics in diffusion models, from easy to hard","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cond-mat.dis-nn","cs.IT","cs.LG","math.IT"],"primary_cat":"stat.ML","authors_text":"Claudia Merger, Lorenzo Bardone, Sebastian Goldt","submitted_at":"2026-03-13T11:07:01Z","abstract_excerpt":"While diffusion models have emerged as a powerful class of generative models, their learning dynamics remain poorly understood. We address this issue first by empirically showing that standard diffusion models trained on natural images exhibit a distributional simplicity bias, learning simple, pair-wise input statistics before specializing to higher-order correlations. We reproduce this behaviour in simple denoisers trained on a minimal data model, the mixed cumulant model, where we precisely control both pair-wise and higher-order correlations of the inputs. We identify a scalar invariant of "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2603.12901","kind":"arxiv","version":1},"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/2603.12901/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":"2603.12901","created_at":"2026-05-21T02:05:00.971335+00:00"},{"alias_kind":"arxiv_version","alias_value":"2603.12901v1","created_at":"2026-05-21T02:05:00.971335+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2603.12901","created_at":"2026-05-21T02:05:00.971335+00:00"},{"alias_kind":"pith_short_12","alias_value":"KLPQSYFTTAOM","created_at":"2026-05-21T02:05:00.971335+00:00"},{"alias_kind":"pith_short_16","alias_value":"KLPQSYFTTAOMX3NQ","created_at":"2026-05-21T02:05:00.971335+00:00"},{"alias_kind":"pith_short_8","alias_value":"KLPQSYFT","created_at":"2026-05-21T02:05:00.971335+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":2,"sample":[{"citing_arxiv_id":"2605.06077","citing_title":"Understanding diffusion models requires rethinking (again) generalization","ref_index":4,"is_internal_anchor":true},{"citing_arxiv_id":"2605.07844","citing_title":"Distributional simplicity bias and effective convexity in Energy Based Models","ref_index":11,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/KLPQSYFTTAOMX3NQD4P5HIUFOY","json":"https://pith.science/pith/KLPQSYFTTAOMX3NQD4P5HIUFOY.json","graph_json":"https://pith.science/api/pith-number/KLPQSYFTTAOMX3NQD4P5HIUFOY/graph.json","events_json":"https://pith.science/api/pith-number/KLPQSYFTTAOMX3NQD4P5HIUFOY/events.json","paper":"https://pith.science/paper/KLPQSYFT"},"agent_actions":{"view_html":"https://pith.science/pith/KLPQSYFTTAOMX3NQD4P5HIUFOY","download_json":"https://pith.science/pith/KLPQSYFTTAOMX3NQD4P5HIUFOY.json","view_paper":"https://pith.science/paper/KLPQSYFT","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2603.12901&json=true","fetch_graph":"https://pith.science/api/pith-number/KLPQSYFTTAOMX3NQD4P5HIUFOY/graph.json","fetch_events":"https://pith.science/api/pith-number/KLPQSYFTTAOMX3NQD4P5HIUFOY/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/KLPQSYFTTAOMX3NQD4P5HIUFOY/action/timestamp_anchor","attest_storage":"https://pith.science/pith/KLPQSYFTTAOMX3NQD4P5HIUFOY/action/storage_attestation","attest_author":"https://pith.science/pith/KLPQSYFTTAOMX3NQD4P5HIUFOY/action/author_attestation","sign_citation":"https://pith.science/pith/KLPQSYFTTAOMX3NQD4P5HIUFOY/action/citation_signature","submit_replication":"https://pith.science/pith/KLPQSYFTTAOMX3NQD4P5HIUFOY/action/replication_record"}},"created_at":"2026-05-21T02:05:00.971335+00:00","updated_at":"2026-05-21T02:05:00.971335+00:00"}