{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:UL7GSGNZZDDOMQRDDIZIJKMAU2","short_pith_number":"pith:UL7GSGNZ","schema_version":"1.0","canonical_sha256":"a2fe6919b9c8c6e642231a3284a980a69962ce910abf2f47335ab30e1b00edcd","source":{"kind":"arxiv","id":"2606.30627","version":1},"attestation_state":"computed","paper":{"title":"Pessimism's Paradox: Conservative Offline Training Amplifies Reward Hacking During Online Adaptation in Reasoning Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","stat.ML"],"primary_cat":"cs.LG","authors_text":"Aman Chadha, Divya Chaudhary, Subramanyam Sahoo, Vinija Jain","submitted_at":"2026-06-29T17:56:03Z","abstract_excerpt":"Conservative offline training is widely advocated as a safe foundation for subsequent online adaptation: if a policy stays close to well-supported behaviour, the argument goes, it is less likely to exploit imperfections in a learned reward model. We challenge this intuition empirically and mechanistically. We train a Qwen3-14B policy under Direct Preference Optimisation (DPO) with three levels of conservatism ($\\beta \\in \\{\\beta_{\\mathrm{lo}}, \\beta_{\\mathrm{mid}}, \\beta_{\\mathrm{hi}}\\}$ derived from empirical log-ratio percentiles), then adapt each checkpoint online against a learned reward e"},"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":"2606.30627","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-06-29T17:56:03Z","cross_cats_sorted":["cs.AI","stat.ML"],"title_canon_sha256":"7367e9c6d242e93e0611c8460ee0bf65a6433389e353b3966ca68fb601491ae9","abstract_canon_sha256":"ee2c30f89e66c211d75b390391b49d43c7ca98b185034cabcac70ba3eb6450c7"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-30T02:18:23.444382Z","signature_b64":"bM7KT9BbUNO0hbI7eciD4UQIj3wa9Uk3uXCYlRvErxhAmD8er/uyaOGZIb8YL5au6Mta5fe30yWEUVzZZF1iBA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"a2fe6919b9c8c6e642231a3284a980a69962ce910abf2f47335ab30e1b00edcd","last_reissued_at":"2026-06-30T02:18:23.443912Z","signature_status":"signed_v1","first_computed_at":"2026-06-30T02:18:23.443912Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Pessimism's Paradox: Conservative Offline Training Amplifies Reward Hacking During Online Adaptation in Reasoning Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","stat.ML"],"primary_cat":"cs.LG","authors_text":"Aman Chadha, Divya Chaudhary, Subramanyam Sahoo, Vinija Jain","submitted_at":"2026-06-29T17:56:03Z","abstract_excerpt":"Conservative offline training is widely advocated as a safe foundation for subsequent online adaptation: if a policy stays close to well-supported behaviour, the argument goes, it is less likely to exploit imperfections in a learned reward model. We challenge this intuition empirically and mechanistically. We train a Qwen3-14B policy under Direct Preference Optimisation (DPO) with three levels of conservatism ($\\beta \\in \\{\\beta_{\\mathrm{lo}}, \\beta_{\\mathrm{mid}}, \\beta_{\\mathrm{hi}}\\}$ derived from empirical log-ratio percentiles), then adapt each checkpoint online against a learned reward e"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2606.30627","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/2606.30627/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":"2606.30627","created_at":"2026-06-30T02:18:23.443970+00:00"},{"alias_kind":"arxiv_version","alias_value":"2606.30627v1","created_at":"2026-06-30T02:18:23.443970+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2606.30627","created_at":"2026-06-30T02:18:23.443970+00:00"},{"alias_kind":"pith_short_12","alias_value":"UL7GSGNZZDDO","created_at":"2026-06-30T02:18:23.443970+00:00"},{"alias_kind":"pith_short_16","alias_value":"UL7GSGNZZDDOMQRD","created_at":"2026-06-30T02:18:23.443970+00:00"},{"alias_kind":"pith_short_8","alias_value":"UL7GSGNZ","created_at":"2026-06-30T02:18:23.443970+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/UL7GSGNZZDDOMQRDDIZIJKMAU2","json":"https://pith.science/pith/UL7GSGNZZDDOMQRDDIZIJKMAU2.json","graph_json":"https://pith.science/api/pith-number/UL7GSGNZZDDOMQRDDIZIJKMAU2/graph.json","events_json":"https://pith.science/api/pith-number/UL7GSGNZZDDOMQRDDIZIJKMAU2/events.json","paper":"https://pith.science/paper/UL7GSGNZ"},"agent_actions":{"view_html":"https://pith.science/pith/UL7GSGNZZDDOMQRDDIZIJKMAU2","download_json":"https://pith.science/pith/UL7GSGNZZDDOMQRDDIZIJKMAU2.json","view_paper":"https://pith.science/paper/UL7GSGNZ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2606.30627&json=true","fetch_graph":"https://pith.science/api/pith-number/UL7GSGNZZDDOMQRDDIZIJKMAU2/graph.json","fetch_events":"https://pith.science/api/pith-number/UL7GSGNZZDDOMQRDDIZIJKMAU2/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/UL7GSGNZZDDOMQRDDIZIJKMAU2/action/timestamp_anchor","attest_storage":"https://pith.science/pith/UL7GSGNZZDDOMQRDDIZIJKMAU2/action/storage_attestation","attest_author":"https://pith.science/pith/UL7GSGNZZDDOMQRDDIZIJKMAU2/action/author_attestation","sign_citation":"https://pith.science/pith/UL7GSGNZZDDOMQRDDIZIJKMAU2/action/citation_signature","submit_replication":"https://pith.science/pith/UL7GSGNZZDDOMQRDDIZIJKMAU2/action/replication_record"}},"created_at":"2026-06-30T02:18:23.443970+00:00","updated_at":"2026-06-30T02:18:23.443970+00:00"}