{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:6CLJL7MIUYKSMJWXZFWJWOIJRW","short_pith_number":"pith:6CLJL7MI","schema_version":"1.0","canonical_sha256":"f09695fd88a6152626d7c96c9b39098d85cb642c1c938589f023589abcc883e8","source":{"kind":"arxiv","id":"2505.24709","version":1},"attestation_state":"computed","paper":{"title":"On Symmetric Losses for Robust Policy Optimization with Noisy Preferences","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Masashi Sugiyama, Soichiro Nishimori, Thanawat Lodkaew, Yu-Jie Zhang","submitted_at":"2025-05-30T15:30:43Z","abstract_excerpt":"Optimizing policies based on human preferences is key to aligning language models with human intent. This work focuses on reward modeling, a core component in reinforcement learning from human feedback (RLHF), and offline preference optimization, such as direct preference optimization. Conventional approaches typically assume accurate annotations. However, real-world preference data often contains noise due to human errors or biases. We propose a principled framework for robust policy optimization under noisy preferences, viewing reward modeling as a classification problem. This allows us to l"},"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":"2505.24709","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-05-30T15:30:43Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"e624563e5b4ce3e306c627fa50a21dade63f9e559675ffe1ff825ab474cab4a4","abstract_canon_sha256":"cbf3d4d2917c5ae60693c7e6c34861252ae0ae20fffdee6a55d8243698a82b63"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:12:55.147559Z","signature_b64":"Hwidcmhcvsq50UD70uS5YcuCBTZnYETtIZO3gPAqGrmR25OV0dnd0K+O7b/06GIxPUQ+Xl2bhrBPFxGzKYreCQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"f09695fd88a6152626d7c96c9b39098d85cb642c1c938589f023589abcc883e8","last_reissued_at":"2026-07-05T11:12:55.147008Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:12:55.147008Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"On Symmetric Losses for Robust Policy Optimization with Noisy Preferences","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Masashi Sugiyama, Soichiro Nishimori, Thanawat Lodkaew, Yu-Jie Zhang","submitted_at":"2025-05-30T15:30:43Z","abstract_excerpt":"Optimizing policies based on human preferences is key to aligning language models with human intent. This work focuses on reward modeling, a core component in reinforcement learning from human feedback (RLHF), and offline preference optimization, such as direct preference optimization. Conventional approaches typically assume accurate annotations. However, real-world preference data often contains noise due to human errors or biases. We propose a principled framework for robust policy optimization under noisy preferences, viewing reward modeling as a classification problem. This allows us to l"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2505.24709","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/2505.24709/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":"2505.24709","created_at":"2026-07-05T11:12:55.147063+00:00"},{"alias_kind":"arxiv_version","alias_value":"2505.24709v1","created_at":"2026-07-05T11:12:55.147063+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2505.24709","created_at":"2026-07-05T11:12:55.147063+00:00"},{"alias_kind":"pith_short_12","alias_value":"6CLJL7MIUYKS","created_at":"2026-07-05T11:12:55.147063+00:00"},{"alias_kind":"pith_short_16","alias_value":"6CLJL7MIUYKSMJWX","created_at":"2026-07-05T11:12:55.147063+00:00"},{"alias_kind":"pith_short_8","alias_value":"6CLJL7MI","created_at":"2026-07-05T11:12:55.147063+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.06036","citing_title":"Optimal Transport for LLM Reward Modeling from Noisy Preference","ref_index":87,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/6CLJL7MIUYKSMJWXZFWJWOIJRW","json":"https://pith.science/pith/6CLJL7MIUYKSMJWXZFWJWOIJRW.json","graph_json":"https://pith.science/api/pith-number/6CLJL7MIUYKSMJWXZFWJWOIJRW/graph.json","events_json":"https://pith.science/api/pith-number/6CLJL7MIUYKSMJWXZFWJWOIJRW/events.json","paper":"https://pith.science/paper/6CLJL7MI"},"agent_actions":{"view_html":"https://pith.science/pith/6CLJL7MIUYKSMJWXZFWJWOIJRW","download_json":"https://pith.science/pith/6CLJL7MIUYKSMJWXZFWJWOIJRW.json","view_paper":"https://pith.science/paper/6CLJL7MI","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2505.24709&json=true","fetch_graph":"https://pith.science/api/pith-number/6CLJL7MIUYKSMJWXZFWJWOIJRW/graph.json","fetch_events":"https://pith.science/api/pith-number/6CLJL7MIUYKSMJWXZFWJWOIJRW/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/6CLJL7MIUYKSMJWXZFWJWOIJRW/action/timestamp_anchor","attest_storage":"https://pith.science/pith/6CLJL7MIUYKSMJWXZFWJWOIJRW/action/storage_attestation","attest_author":"https://pith.science/pith/6CLJL7MIUYKSMJWXZFWJWOIJRW/action/author_attestation","sign_citation":"https://pith.science/pith/6CLJL7MIUYKSMJWXZFWJWOIJRW/action/citation_signature","submit_replication":"https://pith.science/pith/6CLJL7MIUYKSMJWXZFWJWOIJRW/action/replication_record"}},"created_at":"2026-07-05T11:12:55.147063+00:00","updated_at":"2026-07-05T11:12:55.147063+00:00"}