{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2023:P6PJJSPYEB63JEFUKIAOHNCPND","short_pith_number":"pith:P6PJJSPY","schema_version":"1.0","canonical_sha256":"7f9e94c9f8207db490b45200e3b44f68fb28d417c8ebcf69936c99411b029fc3","source":{"kind":"arxiv","id":"2302.03025","version":2},"attestation_state":"computed","paper":{"title":"A Toy Model of Universality: Reverse Engineering How Networks Learn Group Operations","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","math.RT"],"primary_cat":"cs.LG","authors_text":"Bilal Chughtai, Lawrence Chan, Neel Nanda","submitted_at":"2023-02-06T18:59:20Z","abstract_excerpt":"Universality is a key hypothesis in mechanistic interpretability -- that different models learn similar features and circuits when trained on similar tasks. In this work, we study the universality hypothesis by examining how small neural networks learn to implement group composition. We present a novel algorithm by which neural networks may implement composition for any finite group via mathematical representation theory. We then show that networks consistently learn this algorithm by reverse engineering model logits and weights, and confirm our understanding using ablations. By studying netwo"},"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":"2302.03025","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2023-02-06T18:59:20Z","cross_cats_sorted":["cs.AI","math.RT"],"title_canon_sha256":"ea39b044c70a48eb9c09740bf326a9d9566541d895bb644205f921a606682aec","abstract_canon_sha256":"ffa7efdbdfba3bce5c8ecd134ba7e142aa2960367840cfa81d77fadaa758380c"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T06:13:38.317433Z","signature_b64":"RPke0GZRqy88qoqeAQzvmCqbWGTSCPLP6gOhyYvkgIgEdO69wAs4mhjFMntvGb3MNoxrSE+G6Qo1wiOVESCaAg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"7f9e94c9f8207db490b45200e3b44f68fb28d417c8ebcf69936c99411b029fc3","last_reissued_at":"2026-07-05T06:13:38.316814Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T06:13:38.316814Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"A Toy Model of Universality: Reverse Engineering How Networks Learn Group Operations","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","math.RT"],"primary_cat":"cs.LG","authors_text":"Bilal Chughtai, Lawrence Chan, Neel Nanda","submitted_at":"2023-02-06T18:59:20Z","abstract_excerpt":"Universality is a key hypothesis in mechanistic interpretability -- that different models learn similar features and circuits when trained on similar tasks. In this work, we study the universality hypothesis by examining how small neural networks learn to implement group composition. We present a novel algorithm by which neural networks may implement composition for any finite group via mathematical representation theory. We then show that networks consistently learn this algorithm by reverse engineering model logits and weights, and confirm our understanding using ablations. By studying netwo"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2302.03025","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/2302.03025/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":"2302.03025","created_at":"2026-07-05T06:13:38.316920+00:00"},{"alias_kind":"arxiv_version","alias_value":"2302.03025v2","created_at":"2026-07-05T06:13:38.316920+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2302.03025","created_at":"2026-07-05T06:13:38.316920+00:00"},{"alias_kind":"pith_short_12","alias_value":"P6PJJSPYEB63","created_at":"2026-07-05T06:13:38.316920+00:00"},{"alias_kind":"pith_short_16","alias_value":"P6PJJSPYEB63JEFU","created_at":"2026-07-05T06:13:38.316920+00:00"},{"alias_kind":"pith_short_8","alias_value":"P6PJJSPY","created_at":"2026-07-05T06:13:38.316920+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2502.12717","citing_title":"Learning the symmetric group: large from small","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2604.13082","citing_title":"The Long Delay to Arithmetic Generalization: When Learned Representations Outrun Behavior","ref_index":8,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12809","citing_title":"Correcting Influence: Unboxing LLM Outputs with Orthogonal Latent Spaces","ref_index":108,"is_internal_anchor":false},{"citing_arxiv_id":"2605.07407","citing_title":"Emergent Symbolic Structure in Health Foundation Models: Extraction, Alignment, and Cross-Modal Transfer","ref_index":9,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/P6PJJSPYEB63JEFUKIAOHNCPND","json":"https://pith.science/pith/P6PJJSPYEB63JEFUKIAOHNCPND.json","graph_json":"https://pith.science/api/pith-number/P6PJJSPYEB63JEFUKIAOHNCPND/graph.json","events_json":"https://pith.science/api/pith-number/P6PJJSPYEB63JEFUKIAOHNCPND/events.json","paper":"https://pith.science/paper/P6PJJSPY"},"agent_actions":{"view_html":"https://pith.science/pith/P6PJJSPYEB63JEFUKIAOHNCPND","download_json":"https://pith.science/pith/P6PJJSPYEB63JEFUKIAOHNCPND.json","view_paper":"https://pith.science/paper/P6PJJSPY","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2302.03025&json=true","fetch_graph":"https://pith.science/api/pith-number/P6PJJSPYEB63JEFUKIAOHNCPND/graph.json","fetch_events":"https://pith.science/api/pith-number/P6PJJSPYEB63JEFUKIAOHNCPND/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/P6PJJSPYEB63JEFUKIAOHNCPND/action/timestamp_anchor","attest_storage":"https://pith.science/pith/P6PJJSPYEB63JEFUKIAOHNCPND/action/storage_attestation","attest_author":"https://pith.science/pith/P6PJJSPYEB63JEFUKIAOHNCPND/action/author_attestation","sign_citation":"https://pith.science/pith/P6PJJSPYEB63JEFUKIAOHNCPND/action/citation_signature","submit_replication":"https://pith.science/pith/P6PJJSPYEB63JEFUKIAOHNCPND/action/replication_record"}},"created_at":"2026-07-05T06:13:38.316920+00:00","updated_at":"2026-07-05T06:13:38.316920+00:00"}