{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:3CF6TPWUXOC2BJKUVWBF2JKX26","short_pith_number":"pith:3CF6TPWU","schema_version":"1.0","canonical_sha256":"d88be9bed4bb85a0a554ad825d2557d7a8cd03f7ac0ff01b41054b5386f7566c","source":{"kind":"arxiv","id":"2604.00001","version":2},"attestation_state":"computed","paper":{"title":"Filter-then-Weight: Online Data Selection and Reweighting for LLM Fine-Tuning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"An optimizer-aware Filter-then-Weight method improves convergence in online LLM fine-tuning by matching updates to the current optimizer state.","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Fangxin Wang, Henry Peng Zou, Langzhou He, Peyman Baghershahi, Philip S. Yu, Sourav Medya","submitted_at":"2026-03-08T21:46:16Z","abstract_excerpt":"Gradient-based data selection offers a principled framework for estimating sample utility in large language model (LLM) fine-tuning, but existing methods are mostly designed for offline settings. They are therefore less suited to online fine-tuning, where data arrives sequentially, sample utility is step-dependent, and the effective update geometry is shaped by adaptive optimizers. We propose an optimizer-aware framework for gradient-based online data selection and reweighting in LLM fine-tuning. Our key idea is to view online selection not as static sample ranking, but as shaping the next tar"},"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":false},"canonical_record":{"source":{"id":"2604.00001","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-03-08T21:46:16Z","cross_cats_sorted":["cs.AI","cs.CL"],"title_canon_sha256":"1b5148e548e6b4015b9c36280d09946883ff2a59ef212dc0770c29192d5c4b8c","abstract_canon_sha256":"f90620ab9f73f6ef880cec6e7db56e159c3020357f0c959f11f48893daa59db5"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-18T03:10:03.369134Z","signature_b64":"cH7CW083FqlCaSVccpUS3lMAzoIk/cRgtmnd4oGwpu3D6rzbTN9pOQ9A6PcSqrp7X0wpLBKbsWm9IZjjddvTBA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"d88be9bed4bb85a0a554ad825d2557d7a8cd03f7ac0ff01b41054b5386f7566c","last_reissued_at":"2026-05-18T03:10:03.368690Z","signature_status":"signed_v1","first_computed_at":"2026-05-18T03:10:03.368690Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Filter-then-Weight: Online Data Selection and Reweighting for LLM Fine-Tuning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"An optimizer-aware Filter-then-Weight method improves convergence in online LLM fine-tuning by matching updates to the current optimizer state.","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Fangxin Wang, Henry Peng Zou, Langzhou He, Peyman Baghershahi, Philip S. Yu, Sourav Medya","submitted_at":"2026-03-08T21:46:16Z","abstract_excerpt":"Gradient-based data selection offers a principled framework for estimating sample utility in large language model (LLM) fine-tuning, but existing methods are mostly designed for offline settings. They are therefore less suited to online fine-tuning, where data arrives sequentially, sample utility is step-dependent, and the effective update geometry is shaped by adaptive optimizers. We propose an optimizer-aware framework for gradient-based online data selection and reweighting in LLM fine-tuning. Our key idea is to view online selection not as static sample ranking, but as shaping the next tar"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Experiments show that our method consistently improves convergence and downstream performance over existing online data selection baselines under the same data budget.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the optimizer-aware update-matching formulation correctly captures sample utility and that the two-stage filter-plus-weight procedure can be computed efficiently without introducing new biases for long-context LLM data.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Filter-then-Weight is a two-stage optimizer-aware method that filters geometrically useful data candidates and optimizes their coefficients to shape target updates in online LLM fine-tuning.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"An optimizer-aware Filter-then-Weight method improves convergence in online LLM fine-tuning by matching updates to the current optimizer state.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"fa3fc5ae443ef7236f6e1a4aafa1e60424be077bd397cd9b0c862e1aa4bd361f"},"source":{"id":"2604.00001","kind":"arxiv","version":2},"verdict":{"id":"317d1bc6-9c6c-4ad9-a3b3-8123976d230e","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T14:20:09.700915Z","strongest_claim":"Experiments show that our method consistently improves convergence and downstream performance over existing online data selection baselines under the same data budget.","one_line_summary":"Filter-then-Weight is a two-stage optimizer-aware method that filters geometrically useful data candidates and optimizes their coefficients to shape target updates in online LLM fine-tuning.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the optimizer-aware update-matching formulation correctly captures sample utility and that the two-stage filter-plus-weight procedure can be computed efficiently without introducing new biases for long-context LLM data.","pith_extraction_headline":"An optimizer-aware Filter-then-Weight method improves convergence in online LLM fine-tuning by matching updates to the current optimizer state."},"references":{"count":5,"sample":[{"doi":"","year":null,"title":"arXiv preprint arXiv:2308.03296 , year=","work_id":"0d856ab3-d64d-4427-918f-344d35f6818d","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"From quantity to quality: Boosting llm performance with self-guided data selection for instruction tuning","work_id":"6dd32f47-dcfa-4936-8863-f75c6d55383f","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2020,"title":"Yulei Qin, Yuncheng Yang, Pengcheng Guo, Gang Li, Hang Shao, Yuchen Shi, Zihan Xu, Yun Gu, Ke Li, and Xing Sun","work_id":"75dbb3fd-a3d3-4dde-8bcc-72d604931ff9","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Tagcos: Task- agnostic gradient clustered coreset selection for instruction tuning data","work_id":"a91d8610-d88d-479a-bab1-63ec127401b0","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2020,"title":"As discussed in Section 4.4, the top-k filtering based can achieve comparable peak performance, with little increased cost compared with filter-only methods","work_id":"98257d07-0500-4fc1-af48-6da2e2073f50","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":5,"snapshot_sha256":"3f124652fb96add356163ceb12955937946ee8568aa1119ef1faf7beea98c1f7","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":"2604.00001","created_at":"2026-05-18T03:10:03.368752+00:00"},{"alias_kind":"arxiv_version","alias_value":"2604.00001v2","created_at":"2026-05-18T03:10:03.368752+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2604.00001","created_at":"2026-05-18T03:10:03.368752+00:00"},{"alias_kind":"pith_short_12","alias_value":"3CF6TPWUXOC2","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"3CF6TPWUXOC2BJKU","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"3CF6TPWU","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2605.09404","citing_title":"Let the Target Select for Itself: Data Selection via Target-Aligned Paths","ref_index":43,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/3CF6TPWUXOC2BJKUVWBF2JKX26","json":"https://pith.science/pith/3CF6TPWUXOC2BJKUVWBF2JKX26.json","graph_json":"https://pith.science/api/pith-number/3CF6TPWUXOC2BJKUVWBF2JKX26/graph.json","events_json":"https://pith.science/api/pith-number/3CF6TPWUXOC2BJKUVWBF2JKX26/events.json","paper":"https://pith.science/paper/3CF6TPWU"},"agent_actions":{"view_html":"https://pith.science/pith/3CF6TPWUXOC2BJKUVWBF2JKX26","download_json":"https://pith.science/pith/3CF6TPWUXOC2BJKUVWBF2JKX26.json","view_paper":"https://pith.science/paper/3CF6TPWU","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2604.00001&json=true","fetch_graph":"https://pith.science/api/pith-number/3CF6TPWUXOC2BJKUVWBF2JKX26/graph.json","fetch_events":"https://pith.science/api/pith-number/3CF6TPWUXOC2BJKUVWBF2JKX26/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/3CF6TPWUXOC2BJKUVWBF2JKX26/action/timestamp_anchor","attest_storage":"https://pith.science/pith/3CF6TPWUXOC2BJKUVWBF2JKX26/action/storage_attestation","attest_author":"https://pith.science/pith/3CF6TPWUXOC2BJKUVWBF2JKX26/action/author_attestation","sign_citation":"https://pith.science/pith/3CF6TPWUXOC2BJKUVWBF2JKX26/action/citation_signature","submit_replication":"https://pith.science/pith/3CF6TPWUXOC2BJKUVWBF2JKX26/action/replication_record"}},"created_at":"2026-05-18T03:10:03.368752+00:00","updated_at":"2026-05-18T03:10:03.368752+00:00"}