{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:ZLQBVEFR4L5UZF77PLZSQCPPFT","short_pith_number":"pith:ZLQBVEFR","schema_version":"1.0","canonical_sha256":"cae01a90b1e2fb4c97ff7af32809ef2cddd5d00b9a5efed27738091235622f79","source":{"kind":"arxiv","id":"2605.24879","version":1},"attestation_state":"computed","paper":{"title":"Efficient DP-SGD for LLMs with Randomized Clipping","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["math.OC"],"primary_cat":"cs.LG","authors_text":"Devansh Gupta, Enayat Ullah, Huanyu Zhang, Meisam Razaviyayn, Sai Aparna Aketi","submitted_at":"2026-05-24T05:44:20Z","abstract_excerpt":"Large language models (LLMs) are trained on vast datasets that may contain sensitive information. Differential privacy (DP), the de facto standard for formal privacy guarantees, provides a principled framework for training LLMs with provable privacy protection. However, state-of-the-art DP training implementations rely on fast gradient clipping techniques with memory overhead $O(B \\min\\{T^2, d^2\\})$, where $B$ is the batch size, $T$ is the sequence length, and $d$ is the model width. This becomes prohibitive as both model size and context length grow. We propose DP-SGD-RC, a novel variant of D"},"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":"2605.24879","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-05-24T05:44:20Z","cross_cats_sorted":["math.OC"],"title_canon_sha256":"b87713781cf3e80088cb8cd72d21eda8e924e1919d25d54a9855b3976c13d6a3","abstract_canon_sha256":"97cda43c0a1220d30ec59a9ce29938dfb85f6349abd8091576c91419c7f8691b"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-26T01:04:03.163731Z","signature_b64":"9KgnkVmMCTIhH6W0Sd4u2qd6ijg3QZ8sLF3u9tnbRrrmw9WwYWRHkmAoQeqsHMCSGjNdjH+lPM6jDcBzmyOXCg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"cae01a90b1e2fb4c97ff7af32809ef2cddd5d00b9a5efed27738091235622f79","last_reissued_at":"2026-05-26T01:04:03.162882Z","signature_status":"signed_v1","first_computed_at":"2026-05-26T01:04:03.162882Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Efficient DP-SGD for LLMs with Randomized Clipping","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["math.OC"],"primary_cat":"cs.LG","authors_text":"Devansh Gupta, Enayat Ullah, Huanyu Zhang, Meisam Razaviyayn, Sai Aparna Aketi","submitted_at":"2026-05-24T05:44:20Z","abstract_excerpt":"Large language models (LLMs) are trained on vast datasets that may contain sensitive information. Differential privacy (DP), the de facto standard for formal privacy guarantees, provides a principled framework for training LLMs with provable privacy protection. However, state-of-the-art DP training implementations rely on fast gradient clipping techniques with memory overhead $O(B \\min\\{T^2, d^2\\})$, where $B$ is the batch size, $T$ is the sequence length, and $d$ is the model width. This becomes prohibitive as both model size and context length grow. We propose DP-SGD-RC, a novel variant of D"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2605.24879","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/2605.24879/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":"2605.24879","created_at":"2026-05-26T01:04:03.163027+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.24879v1","created_at":"2026-05-26T01:04:03.163027+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.24879","created_at":"2026-05-26T01:04:03.163027+00:00"},{"alias_kind":"pith_short_12","alias_value":"ZLQBVEFR4L5U","created_at":"2026-05-26T01:04:03.163027+00:00"},{"alias_kind":"pith_short_16","alias_value":"ZLQBVEFR4L5UZF77","created_at":"2026-05-26T01:04:03.163027+00:00"},{"alias_kind":"pith_short_8","alias_value":"ZLQBVEFR","created_at":"2026-05-26T01:04:03.163027+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/ZLQBVEFR4L5UZF77PLZSQCPPFT","json":"https://pith.science/pith/ZLQBVEFR4L5UZF77PLZSQCPPFT.json","graph_json":"https://pith.science/api/pith-number/ZLQBVEFR4L5UZF77PLZSQCPPFT/graph.json","events_json":"https://pith.science/api/pith-number/ZLQBVEFR4L5UZF77PLZSQCPPFT/events.json","paper":"https://pith.science/paper/ZLQBVEFR"},"agent_actions":{"view_html":"https://pith.science/pith/ZLQBVEFR4L5UZF77PLZSQCPPFT","download_json":"https://pith.science/pith/ZLQBVEFR4L5UZF77PLZSQCPPFT.json","view_paper":"https://pith.science/paper/ZLQBVEFR","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.24879&json=true","fetch_graph":"https://pith.science/api/pith-number/ZLQBVEFR4L5UZF77PLZSQCPPFT/graph.json","fetch_events":"https://pith.science/api/pith-number/ZLQBVEFR4L5UZF77PLZSQCPPFT/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/ZLQBVEFR4L5UZF77PLZSQCPPFT/action/timestamp_anchor","attest_storage":"https://pith.science/pith/ZLQBVEFR4L5UZF77PLZSQCPPFT/action/storage_attestation","attest_author":"https://pith.science/pith/ZLQBVEFR4L5UZF77PLZSQCPPFT/action/author_attestation","sign_citation":"https://pith.science/pith/ZLQBVEFR4L5UZF77PLZSQCPPFT/action/citation_signature","submit_replication":"https://pith.science/pith/ZLQBVEFR4L5UZF77PLZSQCPPFT/action/replication_record"}},"created_at":"2026-05-26T01:04:03.163027+00:00","updated_at":"2026-05-26T01:04:03.163027+00:00"}