{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:AG42B6F2F4RKMEF3CF7NLVELU2","short_pith_number":"pith:AG42B6F2","schema_version":"1.0","canonical_sha256":"01b9a0f8ba2f22a610bb117ed5d48ba6b486d5a6f938d52cdcbe4a46c7147852","source":{"kind":"arxiv","id":"2502.14289","version":3},"attestation_state":"computed","paper":{"title":"Drift: Decoding-time Personalized Alignments with Implicit User Preferences","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Hwaran Lee, Kang-il Lee, Kyomin Jung, Minbeom Kim, Seongho Joo, Thibaut Thonet","submitted_at":"2025-02-20T06:05:29Z","abstract_excerpt":"Personalized alignments for individual users have been a long-standing goal in large language models (LLMs). We introduce Drift, a novel framework that personalizes LLMs at decoding time with implicit user preferences. Traditional Reinforcement Learning from Human Feedback (RLHF) requires thousands of annotated examples and expensive gradient updates. In contrast, Drift personalizes LLMs in a training-free manner, using only a few dozen examples to steer a frozen model through efficient preference modeling. Our approach models user preferences as a composition of predefined, interpretable attr"},"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":"2502.14289","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2025-02-20T06:05:29Z","cross_cats_sorted":[],"title_canon_sha256":"7c85e66dbed6ef47f06f10e6000fab8800d78a3f42f2e9c12e58eb14b698192d","abstract_canon_sha256":"a4d6f91d75a1507ff139cd58d3de1d1e2354d9cae8936e70655aa3014bc9836b"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:00:04.839177Z","signature_b64":"cCetOf9feJL3nmpmgg2V3Moaxtq7ooPuLBEX3W2GwBuLHl3KmfPR81QcnyNPr8JJ47+0DO+v6ciYscPRAkMJDQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"01b9a0f8ba2f22a610bb117ed5d48ba6b486d5a6f938d52cdcbe4a46c7147852","last_reissued_at":"2026-07-05T11:00:04.838684Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:00:04.838684Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Drift: Decoding-time Personalized Alignments with Implicit User Preferences","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Hwaran Lee, Kang-il Lee, Kyomin Jung, Minbeom Kim, Seongho Joo, Thibaut Thonet","submitted_at":"2025-02-20T06:05:29Z","abstract_excerpt":"Personalized alignments for individual users have been a long-standing goal in large language models (LLMs). We introduce Drift, a novel framework that personalizes LLMs at decoding time with implicit user preferences. Traditional Reinforcement Learning from Human Feedback (RLHF) requires thousands of annotated examples and expensive gradient updates. In contrast, Drift personalizes LLMs in a training-free manner, using only a few dozen examples to steer a frozen model through efficient preference modeling. Our approach models user preferences as a composition of predefined, interpretable attr"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2502.14289","kind":"arxiv","version":3},"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/2502.14289/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":"2502.14289","created_at":"2026-07-05T11:00:04.838742+00:00"},{"alias_kind":"arxiv_version","alias_value":"2502.14289v3","created_at":"2026-07-05T11:00:04.838742+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2502.14289","created_at":"2026-07-05T11:00:04.838742+00:00"},{"alias_kind":"pith_short_12","alias_value":"AG42B6F2F4RK","created_at":"2026-07-05T11:00:04.838742+00:00"},{"alias_kind":"pith_short_16","alias_value":"AG42B6F2F4RKMEF3","created_at":"2026-07-05T11:00:04.838742+00:00"},{"alias_kind":"pith_short_8","alias_value":"AG42B6F2","created_at":"2026-07-05T11:00:04.838742+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.30339","citing_title":"REAR: Test-time Preference Realignment through Reward Decomposition","ref_index":117,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/AG42B6F2F4RKMEF3CF7NLVELU2","json":"https://pith.science/pith/AG42B6F2F4RKMEF3CF7NLVELU2.json","graph_json":"https://pith.science/api/pith-number/AG42B6F2F4RKMEF3CF7NLVELU2/graph.json","events_json":"https://pith.science/api/pith-number/AG42B6F2F4RKMEF3CF7NLVELU2/events.json","paper":"https://pith.science/paper/AG42B6F2"},"agent_actions":{"view_html":"https://pith.science/pith/AG42B6F2F4RKMEF3CF7NLVELU2","download_json":"https://pith.science/pith/AG42B6F2F4RKMEF3CF7NLVELU2.json","view_paper":"https://pith.science/paper/AG42B6F2","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2502.14289&json=true","fetch_graph":"https://pith.science/api/pith-number/AG42B6F2F4RKMEF3CF7NLVELU2/graph.json","fetch_events":"https://pith.science/api/pith-number/AG42B6F2F4RKMEF3CF7NLVELU2/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/AG42B6F2F4RKMEF3CF7NLVELU2/action/timestamp_anchor","attest_storage":"https://pith.science/pith/AG42B6F2F4RKMEF3CF7NLVELU2/action/storage_attestation","attest_author":"https://pith.science/pith/AG42B6F2F4RKMEF3CF7NLVELU2/action/author_attestation","sign_citation":"https://pith.science/pith/AG42B6F2F4RKMEF3CF7NLVELU2/action/citation_signature","submit_replication":"https://pith.science/pith/AG42B6F2F4RKMEF3CF7NLVELU2/action/replication_record"}},"created_at":"2026-07-05T11:00:04.838742+00:00","updated_at":"2026-07-05T11:00:04.838742+00:00"}