{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:FCVT3PGQWPQE6TUORDWKHQRGOR","short_pith_number":"pith:FCVT3PGQ","schema_version":"1.0","canonical_sha256":"28ab3dbcd0b3e04f4e8e88eca3c226747cc5d4b8663698c56a631bb18aaf620a","source":{"kind":"arxiv","id":"2606.09073","version":1},"attestation_state":"computed","paper":{"title":"A Unifying Lens on Reward Uncertainty in RLHF","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Ely Hahami, Jack Benarroch Jedlicki, Ray Zhou, Yoel Zimmermann","submitted_at":"2026-06-08T06:15:30Z","abstract_excerpt":"Reinforcement learning from human feedback (RLHF) is bottlenecked by \\emph{reward hacking}, where the policy exploits errors in a proxy reward model (RM) and produces high RM scores without genuine quality gains. A natural mitigation is \\emph{pessimism}: penalizing rewards in regions where the RM is uncertain. However, standard scalar RMs provide no principled notion of uncertainty. We argue that the right object is a \\emph{distributional} reward model $p(r\\mid x,y)$. Under either a Bayesian inference or a KL-distributionally robust optimization (KL-DRO) lens, the KL-regularized RLHF objective"},"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":"2606.09073","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-06-08T06:15:30Z","cross_cats_sorted":["cs.AI","cs.CL"],"title_canon_sha256":"451827b14a685e8356f6272542177785b23095c6ef77b1b4a8df41aed4d801aa","abstract_canon_sha256":"139ce02a7345aad79683aba8ceaefe81cc4896eac113a4d97f191876776fb5c2"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-09T02:07:57.176823Z","signature_b64":"NP2sKM2/gD1R+qRoi7XunL5No3kDoBhYsOVdR51RN93hanrimYoI3K25QdRTz8L3X3+1rkaVtvQXDO/ZJbj0Bw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"28ab3dbcd0b3e04f4e8e88eca3c226747cc5d4b8663698c56a631bb18aaf620a","last_reissued_at":"2026-06-09T02:07:57.176264Z","signature_status":"signed_v1","first_computed_at":"2026-06-09T02:07:57.176264Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"A Unifying Lens on Reward Uncertainty in RLHF","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Ely Hahami, Jack Benarroch Jedlicki, Ray Zhou, Yoel Zimmermann","submitted_at":"2026-06-08T06:15:30Z","abstract_excerpt":"Reinforcement learning from human feedback (RLHF) is bottlenecked by \\emph{reward hacking}, where the policy exploits errors in a proxy reward model (RM) and produces high RM scores without genuine quality gains. A natural mitigation is \\emph{pessimism}: penalizing rewards in regions where the RM is uncertain. However, standard scalar RMs provide no principled notion of uncertainty. We argue that the right object is a \\emph{distributional} reward model $p(r\\mid x,y)$. Under either a Bayesian inference or a KL-distributionally robust optimization (KL-DRO) lens, the KL-regularized RLHF objective"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2606.09073","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/2606.09073/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":"2606.09073","created_at":"2026-06-09T02:07:57.176348+00:00"},{"alias_kind":"arxiv_version","alias_value":"2606.09073v1","created_at":"2026-06-09T02:07:57.176348+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2606.09073","created_at":"2026-06-09T02:07:57.176348+00:00"},{"alias_kind":"pith_short_12","alias_value":"FCVT3PGQWPQE","created_at":"2026-06-09T02:07:57.176348+00:00"},{"alias_kind":"pith_short_16","alias_value":"FCVT3PGQWPQE6TUO","created_at":"2026-06-09T02:07:57.176348+00:00"},{"alias_kind":"pith_short_8","alias_value":"FCVT3PGQ","created_at":"2026-06-09T02:07:57.176348+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/FCVT3PGQWPQE6TUORDWKHQRGOR","json":"https://pith.science/pith/FCVT3PGQWPQE6TUORDWKHQRGOR.json","graph_json":"https://pith.science/api/pith-number/FCVT3PGQWPQE6TUORDWKHQRGOR/graph.json","events_json":"https://pith.science/api/pith-number/FCVT3PGQWPQE6TUORDWKHQRGOR/events.json","paper":"https://pith.science/paper/FCVT3PGQ"},"agent_actions":{"view_html":"https://pith.science/pith/FCVT3PGQWPQE6TUORDWKHQRGOR","download_json":"https://pith.science/pith/FCVT3PGQWPQE6TUORDWKHQRGOR.json","view_paper":"https://pith.science/paper/FCVT3PGQ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2606.09073&json=true","fetch_graph":"https://pith.science/api/pith-number/FCVT3PGQWPQE6TUORDWKHQRGOR/graph.json","fetch_events":"https://pith.science/api/pith-number/FCVT3PGQWPQE6TUORDWKHQRGOR/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/FCVT3PGQWPQE6TUORDWKHQRGOR/action/timestamp_anchor","attest_storage":"https://pith.science/pith/FCVT3PGQWPQE6TUORDWKHQRGOR/action/storage_attestation","attest_author":"https://pith.science/pith/FCVT3PGQWPQE6TUORDWKHQRGOR/action/author_attestation","sign_citation":"https://pith.science/pith/FCVT3PGQWPQE6TUORDWKHQRGOR/action/citation_signature","submit_replication":"https://pith.science/pith/FCVT3PGQWPQE6TUORDWKHQRGOR/action/replication_record"}},"created_at":"2026-06-09T02:07:57.176348+00:00","updated_at":"2026-06-09T02:07:57.176348+00:00"}