{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:IUCB4GYJCEEL557S6G53ZUZ6VU","short_pith_number":"pith:IUCB4GYJ","schema_version":"1.0","canonical_sha256":"45041e1b091108bef7f2f1bbbcd33ead39f8612548a91c69d3add56199017bec","source":{"kind":"arxiv","id":"2406.03445","version":1},"attestation_state":"computed","paper":{"title":"Pre-trained Large Language Models Use Fourier Features to Compute Addition","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Deqing Fu, Robin Jia, Tianyi Zhou, Vatsal Sharan","submitted_at":"2024-06-05T16:40:53Z","abstract_excerpt":"Pre-trained large language models (LLMs) exhibit impressive mathematical reasoning capabilities, yet how they compute basic arithmetic, such as addition, remains unclear. This paper shows that pre-trained LLMs add numbers using Fourier features -- dimensions in the hidden state that represent numbers via a set of features sparse in the frequency domain. Within the model, MLP and attention layers use Fourier features in complementary ways: MLP layers primarily approximate the magnitude of the answer using low-frequency features, while attention layers primarily perform modular addition (e.g., c"},"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":"2406.03445","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-06-05T16:40:53Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"bde87c8f622650a94d987079bf7aca8fa9a7576fa3634da50b92c6ec06347011","abstract_canon_sha256":"8a9578c28a643c382d0a4116bef598001664fecfef1c66997c5f5a9d09034f36"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:27:56.825224Z","signature_b64":"8FGAJZBraZY0Dk0wS7tuy1xQIJJsTCIpYngsRxhnIMhdDJk3+teZ9EVanCybm9NCKJnJRUOXDAqVs/qS6XYUDg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"45041e1b091108bef7f2f1bbbcd33ead39f8612548a91c69d3add56199017bec","last_reissued_at":"2026-07-05T08:27:56.824755Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:27:56.824755Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Pre-trained Large Language Models Use Fourier Features to Compute Addition","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Deqing Fu, Robin Jia, Tianyi Zhou, Vatsal Sharan","submitted_at":"2024-06-05T16:40:53Z","abstract_excerpt":"Pre-trained large language models (LLMs) exhibit impressive mathematical reasoning capabilities, yet how they compute basic arithmetic, such as addition, remains unclear. This paper shows that pre-trained LLMs add numbers using Fourier features -- dimensions in the hidden state that represent numbers via a set of features sparse in the frequency domain. Within the model, MLP and attention layers use Fourier features in complementary ways: MLP layers primarily approximate the magnitude of the answer using low-frequency features, while attention layers primarily perform modular addition (e.g., c"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2406.03445","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/2406.03445/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":"2406.03445","created_at":"2026-07-05T08:27:56.824812+00:00"},{"alias_kind":"arxiv_version","alias_value":"2406.03445v1","created_at":"2026-07-05T08:27:56.824812+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2406.03445","created_at":"2026-07-05T08:27:56.824812+00:00"},{"alias_kind":"pith_short_12","alias_value":"IUCB4GYJCEEL","created_at":"2026-07-05T08:27:56.824812+00:00"},{"alias_kind":"pith_short_16","alias_value":"IUCB4GYJCEEL557S","created_at":"2026-07-05T08:27:56.824812+00:00"},{"alias_kind":"pith_short_8","alias_value":"IUCB4GYJ","created_at":"2026-07-05T08:27:56.824812+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2502.09741","citing_title":"FoNE: Precise Single-Token Number Embeddings via Fourier Features","ref_index":49,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/IUCB4GYJCEEL557S6G53ZUZ6VU","json":"https://pith.science/pith/IUCB4GYJCEEL557S6G53ZUZ6VU.json","graph_json":"https://pith.science/api/pith-number/IUCB4GYJCEEL557S6G53ZUZ6VU/graph.json","events_json":"https://pith.science/api/pith-number/IUCB4GYJCEEL557S6G53ZUZ6VU/events.json","paper":"https://pith.science/paper/IUCB4GYJ"},"agent_actions":{"view_html":"https://pith.science/pith/IUCB4GYJCEEL557S6G53ZUZ6VU","download_json":"https://pith.science/pith/IUCB4GYJCEEL557S6G53ZUZ6VU.json","view_paper":"https://pith.science/paper/IUCB4GYJ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2406.03445&json=true","fetch_graph":"https://pith.science/api/pith-number/IUCB4GYJCEEL557S6G53ZUZ6VU/graph.json","fetch_events":"https://pith.science/api/pith-number/IUCB4GYJCEEL557S6G53ZUZ6VU/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/IUCB4GYJCEEL557S6G53ZUZ6VU/action/timestamp_anchor","attest_storage":"https://pith.science/pith/IUCB4GYJCEEL557S6G53ZUZ6VU/action/storage_attestation","attest_author":"https://pith.science/pith/IUCB4GYJCEEL557S6G53ZUZ6VU/action/author_attestation","sign_citation":"https://pith.science/pith/IUCB4GYJCEEL557S6G53ZUZ6VU/action/citation_signature","submit_replication":"https://pith.science/pith/IUCB4GYJCEEL557S6G53ZUZ6VU/action/replication_record"}},"created_at":"2026-07-05T08:27:56.824812+00:00","updated_at":"2026-07-05T08:27:56.824812+00:00"}