{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2023:4GZLISFORWYFPHMHLGN4SDP2P6","short_pith_number":"pith:4GZLISFO","schema_version":"1.0","canonical_sha256":"e1b2b448ae8db0579d87599bc90dfa7f8b70549e054c6ffc140d0ac4dadecf36","source":{"kind":"arxiv","id":"2301.05217","version":3},"attestation_state":"computed","paper":{"title":"Progress measures for grokking via mechanistic interpretability","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Transformers on modular addition learn a Fourier rotation algorithm that gradually replaces memorization during training.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Jacob Steinhardt, Jess Smith, Lawrence Chan, Neel Nanda, Tom Lieberum","submitted_at":"2023-01-12T18:56:49Z","abstract_excerpt":"Neural networks often exhibit emergent behavior, where qualitatively new capabilities arise from scaling up the amount of parameters, training data, or training steps. One approach to understanding emergence is to find continuous \\textit{progress measures} that underlie the seemingly discontinuous qualitative changes. We argue that progress measures can be found via mechanistic interpretability: reverse-engineering learned behaviors into their individual components. As a case study, we investigate the recently-discovered phenomenon of ``grokking'' exhibited by small transformers trained on mod"},"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":true,"formal_links_present":true},"canonical_record":{"source":{"id":"2301.05217","kind":"arxiv","version":3},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2023-01-12T18:56:49Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"2b8da6a68b450756d12923b1c2f434a107b1e93fe063bab567c85bdd1c56c5f9","abstract_canon_sha256":"21d2e0a1f4ee7261ee53ba1f358b8ab53995d1926f8e62a563a00647a1530f9c"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:39:21.567088Z","signature_b64":"zaULf/6h7tzOnYtjRgnZpbgl5sDF/+1oNC78J8yVDW8Os/CBbOuccSizemltAO8enlaUYOOxI8Osn3l4lCooBw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"e1b2b448ae8db0579d87599bc90dfa7f8b70549e054c6ffc140d0ac4dadecf36","last_reissued_at":"2026-05-17T23:39:21.566368Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:39:21.566368Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Progress measures for grokking via mechanistic interpretability","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Transformers on modular addition learn a Fourier rotation algorithm that gradually replaces memorization during training.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Jacob Steinhardt, Jess Smith, Lawrence Chan, Neel Nanda, Tom Lieberum","submitted_at":"2023-01-12T18:56:49Z","abstract_excerpt":"Neural networks often exhibit emergent behavior, where qualitatively new capabilities arise from scaling up the amount of parameters, training data, or training steps. One approach to understanding emergence is to find continuous \\textit{progress measures} that underlie the seemingly discontinuous qualitative changes. We argue that progress measures can be found via mechanistic interpretability: reverse-engineering learned behaviors into their individual components. As a case study, we investigate the recently-discovered phenomenon of ``grokking'' exhibited by small transformers trained on mod"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We fully reverse engineer the algorithm learned by these networks, which uses discrete Fourier transforms and trigonometric identities to convert addition to rotation about a circle. We confirm the algorithm by analyzing the activations and weights and by performing ablations in Fourier space.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the identified Fourier circuit is the dominant mechanism and that ablations in Fourier space fully isolate it without missing other co-occurring computations that could also produce the observed behavior.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Grokking arises from gradual amplification of a Fourier-based circuit in the weights followed by removal of memorizing components.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Transformers on modular addition learn a Fourier rotation algorithm that gradually replaces memorization during training.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"f1565aaea603f2b8e3c5053c34253097b56be15b3dce4f00a30faa1a6f4836b7"},"source":{"id":"2301.05217","kind":"arxiv","version":3},"verdict":{"id":"29c7ba9c-7acf-4806-ba94-b355101b2be4","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T21:47:29.091836Z","strongest_claim":"We fully reverse engineer the algorithm learned by these networks, which uses discrete Fourier transforms and trigonometric identities to convert addition to rotation about a circle. We confirm the algorithm by analyzing the activations and weights and by performing ablations in Fourier space.","one_line_summary":"Grokking arises from gradual amplification of a Fourier-based circuit in the weights followed by removal of memorizing components.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the identified Fourier circuit is the dominant mechanism and that ablations in Fourier space fully isolate it without missing other co-occurring computations that could also produce the observed behavior.","pith_extraction_headline":"Transformers on modular addition learn a Fourier rotation algorithm that gradually replaces memorization during training."},"references":{"count":43,"sample":[{"doi":"","year":null,"title":"More is different for AI , url=","work_id":"b83cee25-39c1-43e8-a9b2-b44d309946a1","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"OpenAI blog , volume=","work_id":"31dc92c3-2fc3-432b-8d63-b7ee13f53a9c","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Advances in neural information processing systems , volume=","work_id":"12f5a236-ef7a-4d13-b4de-b51465a6f977","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"2022 ACM Conference on Fairness, Accountability, and Transparency , pages=","work_id":"13cfa27e-b611-4974-9c7b-cc1f7aad2bfc","ref_index":9,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Beren's Blog - Thoughts on AI, Neuroscience, and other things that interest me","work_id":"8c046681-98d7-4ee8-8d21-c4ab6dd3030c","ref_index":10,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":43,"snapshot_sha256":"a5a08dc1548e305c79acd181969ac690e75c98e831e57bfe9e16026165d8dbe2","internal_anchors":8},"formal_canon":{"evidence_count":3,"snapshot_sha256":"31980ee767ce7529ad6c2d14bea13086d7092d3330d6d0dcde84c0b1b5a67bc6"},"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":"2301.05217","created_at":"2026-05-17T23:39:21.566481+00:00"},{"alias_kind":"arxiv_version","alias_value":"2301.05217v3","created_at":"2026-05-17T23:39:21.566481+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2301.05217","created_at":"2026-05-17T23:39:21.566481+00:00"},{"alias_kind":"pith_short_12","alias_value":"4GZLISFORWYF","created_at":"2026-05-18T12:33:33.725879+00:00"},{"alias_kind":"pith_short_16","alias_value":"4GZLISFORWYFPHMH","created_at":"2026-05-18T12:33:33.725879+00:00"},{"alias_kind":"pith_short_8","alias_value":"4GZLISFO","created_at":"2026-05-18T12:33:33.725879+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":40,"internal_anchor_count":40,"sample":[{"citing_arxiv_id":"2605.23040","citing_title":"Steered Generation via Gradient-Based Optimization on Sparse Query Features","ref_index":31,"is_internal_anchor":true},{"citing_arxiv_id":"2502.09741","citing_title":"FoNE: Precise Single-Token Number Embeddings via Fourier Features","ref_index":29,"is_internal_anchor":true},{"citing_arxiv_id":"2502.12717","citing_title":"Learning the symmetric group: large from small","ref_index":13,"is_internal_anchor":true},{"citing_arxiv_id":"2605.22679","citing_title":"Conceptualizing Embeddings: Sparse Disentanglement for Vision-Language Models","ref_index":12,"is_internal_anchor":true},{"citing_arxiv_id":"2510.04930","citing_title":"Egalitarian Gradient Descent: A Simple Approach to Accelerated Grokking","ref_index":7,"is_internal_anchor":true},{"citing_arxiv_id":"2512.22471","citing_title":"The Bayesian Geometry of Transformer Attention","ref_index":13,"is_internal_anchor":true},{"citing_arxiv_id":"2605.20202","citing_title":"Under Pressure: Emotional Framing Induces Measurable Behavioral Shifts and Structured Internal Geometry in Small Language Models","ref_index":10,"is_internal_anchor":true},{"citing_arxiv_id":"2605.20299","citing_title":"Mechanisms of Misgeneralization in Physical Sequence Modeling","ref_index":61,"is_internal_anchor":true},{"citing_arxiv_id":"2605.20824","citing_title":"Markovian Circuit Tracing for Transformer State Dynamic","ref_index":44,"is_internal_anchor":true},{"citing_arxiv_id":"2605.16325","citing_title":"Phase Transitions in Driven Informational Systems: A Two-Field Perspective on Learning Theory and Non-Equilibrium Chemistry","ref_index":4,"is_internal_anchor":true},{"citing_arxiv_id":"2605.18845","citing_title":"First-Passage Prediction of Grokking Delay: ACalibrated Law under AdamW with Causal Validation","ref_index":3,"is_internal_anchor":true},{"citing_arxiv_id":"2605.03258","citing_title":"The Right Answer, the Wrong Direction: Why Transformers Fail at Counting and How to Fix It","ref_index":7,"is_internal_anchor":true},{"citing_arxiv_id":"2510.00468","citing_title":"Feature Identification via the Empirical NTK","ref_index":8,"is_internal_anchor":true},{"citing_arxiv_id":"2511.02627","citing_title":"DecompSR: A dataset for decomposed analyses of compositional multihop spatial reasoning","ref_index":15,"is_internal_anchor":true},{"citing_arxiv_id":"2601.19208","citing_title":"How Do Transformers Learn to Associate Tokens: Gradient Leading Terms Bring Mechanistic Interpretability","ref_index":14,"is_internal_anchor":true},{"citing_arxiv_id":"2504.20571","citing_title":"Reinforcement Learning for Reasoning in Large Language Models with One Training Example","ref_index":38,"is_internal_anchor":true},{"citing_arxiv_id":"2605.12394","citing_title":"Detecting overfitting in Neural Networks during long-horizon grokking using Random Matrix Theory","ref_index":17,"is_internal_anchor":true},{"citing_arxiv_id":"2605.15183","citing_title":"When Are Two Networks the Same? Tensor Similarity for Mechanistic Interpretability","ref_index":20,"is_internal_anchor":true},{"citing_arxiv_id":"2604.13082","citing_title":"The Long Delay to Arithmetic Generalization: When Learned Representations Outrun Behavior","ref_index":18,"is_internal_anchor":true},{"citing_arxiv_id":"2605.12809","citing_title":"Correcting Influence: Unboxing LLM Outputs with Orthogonal Latent Spaces","ref_index":109,"is_internal_anchor":true},{"citing_arxiv_id":"2605.13450","citing_title":"Assessing the Creativity of Large Language Models: Testing, Limits, and New Frontiers","ref_index":11,"is_internal_anchor":true},{"citing_arxiv_id":"2604.01590","citing_title":"PhiNet: Speaker Verification with Phonetic Interpretability","ref_index":58,"is_internal_anchor":true},{"citing_arxiv_id":"2605.12207","citing_title":"Not How Many, But Which: Parameter Placement in Low-Rank Adaptation","ref_index":65,"is_internal_anchor":true},{"citing_arxiv_id":"2605.12394","citing_title":"Detecting overfitting in Neural Networks during long-horizon grokking using Random Matrix Theory","ref_index":17,"is_internal_anchor":true},{"citing_arxiv_id":"2605.08119","citing_title":"Feature Repulsion and Spectral Lock-in: An Empirical Study of Two-Layer Network Grokking","ref_index":3,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":3,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/4GZLISFORWYFPHMHLGN4SDP2P6","json":"https://pith.science/pith/4GZLISFORWYFPHMHLGN4SDP2P6.json","graph_json":"https://pith.science/api/pith-number/4GZLISFORWYFPHMHLGN4SDP2P6/graph.json","events_json":"https://pith.science/api/pith-number/4GZLISFORWYFPHMHLGN4SDP2P6/events.json","paper":"https://pith.science/paper/4GZLISFO"},"agent_actions":{"view_html":"https://pith.science/pith/4GZLISFORWYFPHMHLGN4SDP2P6","download_json":"https://pith.science/pith/4GZLISFORWYFPHMHLGN4SDP2P6.json","view_paper":"https://pith.science/paper/4GZLISFO","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2301.05217&json=true","fetch_graph":"https://pith.science/api/pith-number/4GZLISFORWYFPHMHLGN4SDP2P6/graph.json","fetch_events":"https://pith.science/api/pith-number/4GZLISFORWYFPHMHLGN4SDP2P6/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/4GZLISFORWYFPHMHLGN4SDP2P6/action/timestamp_anchor","attest_storage":"https://pith.science/pith/4GZLISFORWYFPHMHLGN4SDP2P6/action/storage_attestation","attest_author":"https://pith.science/pith/4GZLISFORWYFPHMHLGN4SDP2P6/action/author_attestation","sign_citation":"https://pith.science/pith/4GZLISFORWYFPHMHLGN4SDP2P6/action/citation_signature","submit_replication":"https://pith.science/pith/4GZLISFORWYFPHMHLGN4SDP2P6/action/replication_record"}},"created_at":"2026-05-17T23:39:21.566481+00:00","updated_at":"2026-05-17T23:39:21.566481+00:00"}