{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2022:J3PUJK7AYXGISUUNRUOZSQWSFY","short_pith_number":"pith:J3PUJK7A","schema_version":"1.0","canonical_sha256":"4edf44abe0c5cc89528d8d1d9942d22e035c725ce8ad0f0be8e5e567a64a9b95","source":{"kind":"arxiv","id":"2206.04817","version":2},"attestation_state":"computed","paper":{"title":"The Slingshot Mechanism: An Empirical Study of Adaptive Optimizers and the Grokking Phenomenon","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["math.OC"],"primary_cat":"cs.LG","authors_text":"Etai Littwin, Joshua Susskind, Omid Saremi, Roni Paiss, Shuangfei Zhai, Vimal Thilak","submitted_at":"2022-06-10T00:04:21Z","abstract_excerpt":"The grokking phenomenon as reported by Power et al. ( arXiv:2201.02177 ) refers to a regime where a long period of overfitting is followed by a seemingly sudden transition to perfect generalization. In this paper, we attempt to reveal the underpinnings of Grokking via a series of empirical studies. Specifically, we uncover an optimization anomaly plaguing adaptive optimizers at extremely late stages of training, referred to as the Slingshot Mechanism. A prominent artifact of the Slingshot Mechanism can be measured by the cyclic phase transitions between stable and unstable training regimes, an"},"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":"2206.04817","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2022-06-10T00:04:21Z","cross_cats_sorted":["math.OC"],"title_canon_sha256":"37c970c61606096712836c57d3898f58c09ed162ccc471f9d30f7aa20bdb3a8c","abstract_canon_sha256":"daa60e7a0cb20168ea9b125d25d7a4fe85de5b69bb32d4dd0fbfadc3a697fb1d"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T04:30:57.506076Z","signature_b64":"A3ZQLlxZbM8e0ldDmetBW2y34ByDyri76/pH0ADUztNK/yurdU4YQNcbB32EinTZvcfS0b/YdiS+fNnRIAozAQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"4edf44abe0c5cc89528d8d1d9942d22e035c725ce8ad0f0be8e5e567a64a9b95","last_reissued_at":"2026-07-05T04:30:57.505509Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T04:30:57.505509Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"The Slingshot Mechanism: An Empirical Study of Adaptive Optimizers and the Grokking Phenomenon","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["math.OC"],"primary_cat":"cs.LG","authors_text":"Etai Littwin, Joshua Susskind, Omid Saremi, Roni Paiss, Shuangfei Zhai, Vimal Thilak","submitted_at":"2022-06-10T00:04:21Z","abstract_excerpt":"The grokking phenomenon as reported by Power et al. ( arXiv:2201.02177 ) refers to a regime where a long period of overfitting is followed by a seemingly sudden transition to perfect generalization. In this paper, we attempt to reveal the underpinnings of Grokking via a series of empirical studies. Specifically, we uncover an optimization anomaly plaguing adaptive optimizers at extremely late stages of training, referred to as the Slingshot Mechanism. A prominent artifact of the Slingshot Mechanism can be measured by the cyclic phase transitions between stable and unstable training regimes, an"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2206.04817","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/2206.04817/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":"2206.04817","created_at":"2026-07-05T04:30:57.505587+00:00"},{"alias_kind":"arxiv_version","alias_value":"2206.04817v2","created_at":"2026-07-05T04:30:57.505587+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2206.04817","created_at":"2026-07-05T04:30:57.505587+00:00"},{"alias_kind":"pith_short_12","alias_value":"J3PUJK7AYXGI","created_at":"2026-07-05T04:30:57.505587+00:00"},{"alias_kind":"pith_short_16","alias_value":"J3PUJK7AYXGISUUN","created_at":"2026-07-05T04:30:57.505587+00:00"},{"alias_kind":"pith_short_8","alias_value":"J3PUJK7A","created_at":"2026-07-05T04:30:57.505587+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":16,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.22873","citing_title":"SingGuard: A Policy-Adaptive Multimodal LLM Guardrail with Dynamic Reasoning","ref_index":71,"is_internal_anchor":false},{"citing_arxiv_id":"2606.12966","citing_title":"Circuit Synchronization Precedes Generalization: A Causal Precursor to Grokking","ref_index":11,"is_internal_anchor":false},{"citing_arxiv_id":"2606.22873","citing_title":"SingGuard: A Policy-Adaptive Multimodal LLM Guardrail with Dynamic Reasoning","ref_index":71,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06152","citing_title":"Grokking or Glitching? How Low-Precision Drives Slingshot Loss Spikes","ref_index":1,"is_internal_anchor":false},{"citing_arxiv_id":"2606.29679","citing_title":"Learning as Observable Matrix Dynamics: Diffusive Relaxations versus Phase Transitions","ref_index":21,"is_internal_anchor":false},{"citing_arxiv_id":"2605.17767","citing_title":"Feature Learning in Linear-Width Two-Layer Networks: Two vs. One Step of Gradient Descent","ref_index":204,"is_internal_anchor":false},{"citing_arxiv_id":"2510.04930","citing_title":"Egalitarian Gradient Descent: A Simple Approach to Accelerated Grokking","ref_index":12,"is_internal_anchor":false},{"citing_arxiv_id":"2605.17767","citing_title":"Feature Learning in Linear-Width Two-Layer Networks: Two vs. One Step of Gradient Descent","ref_index":204,"is_internal_anchor":false},{"citing_arxiv_id":"2601.03162","citing_title":"On the Convergence Behavior of Preconditioned Gradient Descent Toward the Rich Learning Regime","ref_index":18,"is_internal_anchor":false},{"citing_arxiv_id":"2603.13331","citing_title":"The Norm-Separation Delay Law of Grokking: A First-Principles Theory of Delayed Generalization","ref_index":11,"is_internal_anchor":false},{"citing_arxiv_id":"2603.13331","citing_title":"The Norm-Separation Delay Law of Grokking: A First-Principles Theory of Delayed Generalization","ref_index":12,"is_internal_anchor":false},{"citing_arxiv_id":"2301.05217","citing_title":"Progress measures for grokking via mechanistic interpretability","ref_index":51,"is_internal_anchor":false},{"citing_arxiv_id":"2604.04171","citing_title":"A Model of Understanding in Deep Learning Systems","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06152","citing_title":"Grokking or Glitching? How Low-Precision Drives Slingshot Loss Spikes","ref_index":1,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09724","citing_title":"Model Capacity Determines Grokking through Competing Memorisation and Generalisation Speeds","ref_index":34,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06152","citing_title":"Grokking or Glitching? How Low-Precision Drives Slingshot Loss Spikes","ref_index":1,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/J3PUJK7AYXGISUUNRUOZSQWSFY","json":"https://pith.science/pith/J3PUJK7AYXGISUUNRUOZSQWSFY.json","graph_json":"https://pith.science/api/pith-number/J3PUJK7AYXGISUUNRUOZSQWSFY/graph.json","events_json":"https://pith.science/api/pith-number/J3PUJK7AYXGISUUNRUOZSQWSFY/events.json","paper":"https://pith.science/paper/J3PUJK7A"},"agent_actions":{"view_html":"https://pith.science/pith/J3PUJK7AYXGISUUNRUOZSQWSFY","download_json":"https://pith.science/pith/J3PUJK7AYXGISUUNRUOZSQWSFY.json","view_paper":"https://pith.science/paper/J3PUJK7A","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2206.04817&json=true","fetch_graph":"https://pith.science/api/pith-number/J3PUJK7AYXGISUUNRUOZSQWSFY/graph.json","fetch_events":"https://pith.science/api/pith-number/J3PUJK7AYXGISUUNRUOZSQWSFY/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/J3PUJK7AYXGISUUNRUOZSQWSFY/action/timestamp_anchor","attest_storage":"https://pith.science/pith/J3PUJK7AYXGISUUNRUOZSQWSFY/action/storage_attestation","attest_author":"https://pith.science/pith/J3PUJK7AYXGISUUNRUOZSQWSFY/action/author_attestation","sign_citation":"https://pith.science/pith/J3PUJK7AYXGISUUNRUOZSQWSFY/action/citation_signature","submit_replication":"https://pith.science/pith/J3PUJK7AYXGISUUNRUOZSQWSFY/action/replication_record"}},"created_at":"2026-07-05T04:30:57.505587+00:00","updated_at":"2026-07-05T04:30:57.505587+00:00"}