{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:RITIFKOXBVM3H2XGQT4N23QQAZ","short_pith_number":"pith:RITIFKOX","schema_version":"1.0","canonical_sha256":"8a2682a9d70d59b3eae684f8dd6e1006674a20be9c82ad8450a542a6432ce51d","source":{"kind":"arxiv","id":"2602.06239","version":2},"attestation_state":"computed","paper":{"title":"Provably avoiding over-optimization in Direct Preference Optimization without knowing the data distribution","license":"http://creativecommons.org/licenses/by/4.0/","headline":"PEPO mitigates DPO over-optimization by achieving sample complexity bounds that depend only on single-policy concentrability.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Adam Barla, Emanuele Nevali, Luca Viano, Volkan Cevher","submitted_at":"2026-02-05T22:31:07Z","abstract_excerpt":"We introduce PEPO (Pessimistic Ensemble based Preference Optimization), a single-step Direct Preference Optimization (DPO)-like algorithm to mitigate the well-known over-optimization issue in preference learning without requiring the knowledge of the data-generating distribution or learning an explicit reward model. PEPO achieves pessimism via an ensemble of preference-optimized policies trained on disjoint data subsets and then aggregates them through a worst case construction that favors the agreement across models. In the tabular setting, PEPO achieves sample complexity guarantees depending"},"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":"2602.06239","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-02-05T22:31:07Z","cross_cats_sorted":[],"title_canon_sha256":"db7ddacc3944814b6e89500111aa47ddc7ffbdfc4cd5b094ac1b2867786e411b","abstract_canon_sha256":"138b14a500ed947f80d010919a83f4bdc3475151f49ad3e001a915a05d5c0d43"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-18T02:44:31.368909Z","signature_b64":"EFk47imAmxA499dUTgm7++GstpNtyV3oZtBwJDStrXh6w5OLIWWSy6hP4oRGmftKIcv2FbVE6iYCnmUNwfwYCg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"8a2682a9d70d59b3eae684f8dd6e1006674a20be9c82ad8450a542a6432ce51d","last_reissued_at":"2026-05-18T02:44:31.368347Z","signature_status":"signed_v1","first_computed_at":"2026-05-18T02:44:31.368347Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Provably avoiding over-optimization in Direct Preference Optimization without knowing the data distribution","license":"http://creativecommons.org/licenses/by/4.0/","headline":"PEPO mitigates DPO over-optimization by achieving sample complexity bounds that depend only on single-policy concentrability.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Adam Barla, Emanuele Nevali, Luca Viano, Volkan Cevher","submitted_at":"2026-02-05T22:31:07Z","abstract_excerpt":"We introduce PEPO (Pessimistic Ensemble based Preference Optimization), a single-step Direct Preference Optimization (DPO)-like algorithm to mitigate the well-known over-optimization issue in preference learning without requiring the knowledge of the data-generating distribution or learning an explicit reward model. PEPO achieves pessimism via an ensemble of preference-optimized policies trained on disjoint data subsets and then aggregates them through a worst case construction that favors the agreement across models. In the tabular setting, PEPO achieves sample complexity guarantees depending"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"In the tabular setting, PEPO achieves sample complexity guarantees depending only on a single-policy concentrability coefficient, thus avoiding the all-policy concentrability which affects the guarantees of algorithms prone to over-optimization, such as DPO.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That an ensemble of policies trained on disjoint subsets can be aggregated via worst-case construction to produce pessimism without access to the data-generating distribution or explicit reward model.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"PEPO uses pessimistic ensembling of DPO policies on data subsets to achieve single-policy concentrability sample bounds and avoid over-optimization in tabular settings.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"PEPO mitigates DPO over-optimization by achieving sample complexity bounds that depend only on single-policy concentrability.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"9e2f8d8968b7d60ffb3f409dec81c4d8529ae42b5722e3dbf8ae85e7fb8c9da1"},"source":{"id":"2602.06239","kind":"arxiv","version":2},"verdict":{"id":"8e62023a-b6ba-4e51-9bc1-9df721e97474","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T06:32:34.967797Z","strongest_claim":"In the tabular setting, PEPO achieves sample complexity guarantees depending only on a single-policy concentrability coefficient, thus avoiding the all-policy concentrability which affects the guarantees of algorithms prone to over-optimization, such as DPO.","one_line_summary":"PEPO uses pessimistic ensembling of DPO policies on data subsets to achieve single-policy concentrability sample bounds and avoid over-optimization in tabular settings.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That an ensemble of policies trained on disjoint subsets can be aggregated via worst-case construction to produce pessimism without access to the data-generating distribution or explicit reward model.","pith_extraction_headline":"PEPO mitigates DPO over-optimization by achieving sample complexity bounds that depend only on single-policy concentrability."},"references":{"count":49,"sample":[{"doi":"","year":null,"title":"Design considerations in offline preference-based rl","work_id":"09c3c6cb-11d2-44d6-a1ad-296f9a473b37","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"XRPO: Pushing the limits of GRPO with targeted exploration and exploitation","work_id":"c1e825b8-b415-4c32-a6a6-b434dbb9a276","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Value-incentivized preference optimization: A unified approach to online and offline rlhf.arXiv preprint arXiv:2405.19320,","work_id":"6d4c354e-e135-4ac0-8b34-169bdac4b908","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"On extending direct preference optimization to accommodate ties.arXiv preprint arXiv:2409.17431,","work_id":"0068f7ac-2b68-441b-8d52-de280b26eb0e","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"AvoidingO(eRmax)scaling in rlhf through preference-based exploration.arXiv preprint arXiv:2502.00666,","work_id":"3c83224d-7ff1-4379-960c-e55e2b2fe73f","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":49,"snapshot_sha256":"07dc1bd4215a1b76fb1d112528984e69821ad42546df3e7c957605760659cb0b","internal_anchors":5},"formal_canon":{"evidence_count":2,"snapshot_sha256":"dc323d102f41ea093701f3a0eef8044f6c409d8753281be99a33d9fd22e68810"},"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":"2602.06239","created_at":"2026-05-18T02:44:31.368428+00:00"},{"alias_kind":"arxiv_version","alias_value":"2602.06239v2","created_at":"2026-05-18T02:44:31.368428+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2602.06239","created_at":"2026-05-18T02:44:31.368428+00:00"},{"alias_kind":"pith_short_12","alias_value":"RITIFKOXBVM3","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"RITIFKOXBVM3H2XG","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"RITIFKOX","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":2,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/RITIFKOXBVM3H2XGQT4N23QQAZ","json":"https://pith.science/pith/RITIFKOXBVM3H2XGQT4N23QQAZ.json","graph_json":"https://pith.science/api/pith-number/RITIFKOXBVM3H2XGQT4N23QQAZ/graph.json","events_json":"https://pith.science/api/pith-number/RITIFKOXBVM3H2XGQT4N23QQAZ/events.json","paper":"https://pith.science/paper/RITIFKOX"},"agent_actions":{"view_html":"https://pith.science/pith/RITIFKOXBVM3H2XGQT4N23QQAZ","download_json":"https://pith.science/pith/RITIFKOXBVM3H2XGQT4N23QQAZ.json","view_paper":"https://pith.science/paper/RITIFKOX","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2602.06239&json=true","fetch_graph":"https://pith.science/api/pith-number/RITIFKOXBVM3H2XGQT4N23QQAZ/graph.json","fetch_events":"https://pith.science/api/pith-number/RITIFKOXBVM3H2XGQT4N23QQAZ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/RITIFKOXBVM3H2XGQT4N23QQAZ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/RITIFKOXBVM3H2XGQT4N23QQAZ/action/storage_attestation","attest_author":"https://pith.science/pith/RITIFKOXBVM3H2XGQT4N23QQAZ/action/author_attestation","sign_citation":"https://pith.science/pith/RITIFKOXBVM3H2XGQT4N23QQAZ/action/citation_signature","submit_replication":"https://pith.science/pith/RITIFKOXBVM3H2XGQT4N23QQAZ/action/replication_record"}},"created_at":"2026-05-18T02:44:31.368428+00:00","updated_at":"2026-05-18T02:44:31.368428+00:00"}