{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:R6APKS7T34X7NKFOYYN5A6CNVD","short_pith_number":"pith:R6APKS7T","schema_version":"1.0","canonical_sha256":"8f80f54bf3df2ff6a8aec61bd0784da8ef0c8a0f060c2524ed25fcc582aa2e80","source":{"kind":"arxiv","id":"2602.18037","version":2},"attestation_state":"computed","paper":{"title":"Gradient Regularization Mitigates Reward Hacking in Reinforcement Learning from Human Feedback and Verifiable Rewards","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Johannes Ackermann, Masashi Sugiyama, Michael Noukhovitch, Takashi Ishida","submitted_at":"2026-02-20T07:32:22Z","abstract_excerpt":"Reinforcement Learning from Human Feedback (RLHF) or Verifiable Rewards (RLVR) are two key steps in the post-training of modern Language Models (LMs). A common problem is reward hacking, where the policy may exploit inaccuracies of the reward and learn an unintended behavior. Most previous works address this by limiting the policy update with a Kullback-Leibler (KL) penalty towards a reference model. We propose a different framing: Train the LM in a way that biases policy updates towards regions in which the reward is more accurate. First, we derive a theoretical connection between the accurac"},"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":"2602.18037","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-02-20T07:32:22Z","cross_cats_sorted":["cs.AI","cs.CL"],"title_canon_sha256":"016d3b075de16b680a5ccb96eac499c6725bb6a10992f4fc17c241d1eb0fe789","abstract_canon_sha256":"4132ef63275c09cf4a368e95f371a762ba9a20ee2aec859a848c05f2386ddee6"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-07T00:15:52.603871Z","signature_b64":"bWjQan/2HS30Uq+6pwRutjnlaMNyV+x2gPZOP17P7BJv0NROx5fgVqALtHhMS1qBUyYTuiPL6ulkmwHd2FqnDQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"8f80f54bf3df2ff6a8aec61bd0784da8ef0c8a0f060c2524ed25fcc582aa2e80","last_reissued_at":"2026-07-07T00:15:52.603114Z","signature_status":"signed_v1","first_computed_at":"2026-07-07T00:15:52.603114Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Gradient Regularization Mitigates Reward Hacking in Reinforcement Learning from Human Feedback and Verifiable Rewards","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Johannes Ackermann, Masashi Sugiyama, Michael Noukhovitch, Takashi Ishida","submitted_at":"2026-02-20T07:32:22Z","abstract_excerpt":"Reinforcement Learning from Human Feedback (RLHF) or Verifiable Rewards (RLVR) are two key steps in the post-training of modern Language Models (LMs). A common problem is reward hacking, where the policy may exploit inaccuracies of the reward and learn an unintended behavior. Most previous works address this by limiting the policy update with a Kullback-Leibler (KL) penalty towards a reference model. We propose a different framing: Train the LM in a way that biases policy updates towards regions in which the reward is more accurate. First, we derive a theoretical connection between the accurac"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2602.18037","kind":"arxiv","version":2},"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/2602.18037/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":"2602.18037","created_at":"2026-07-07T00:15:52.603232+00:00"},{"alias_kind":"arxiv_version","alias_value":"2602.18037v2","created_at":"2026-07-07T00:15:52.603232+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2602.18037","created_at":"2026-07-07T00:15:52.603232+00:00"},{"alias_kind":"pith_short_12","alias_value":"R6APKS7T34X7","created_at":"2026-07-07T00:15:52.603232+00:00"},{"alias_kind":"pith_short_16","alias_value":"R6APKS7T34X7NKFO","created_at":"2026-07-07T00:15:52.603232+00:00"},{"alias_kind":"pith_short_8","alias_value":"R6APKS7T","created_at":"2026-07-07T00:15:52.603232+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":2,"sample":[{"citing_arxiv_id":"2606.25527","citing_title":"Beyond One-Size-Fits-All: Diagnosis-Driven Online Reinforcement Learning with Offline Priors","ref_index":1,"is_internal_anchor":true},{"citing_arxiv_id":"2605.25189","citing_title":"Directional Alignment Mitigates Reward Hacking in Reinforcement Learning for Language Models","ref_index":1,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/R6APKS7T34X7NKFOYYN5A6CNVD","json":"https://pith.science/pith/R6APKS7T34X7NKFOYYN5A6CNVD.json","graph_json":"https://pith.science/api/pith-number/R6APKS7T34X7NKFOYYN5A6CNVD/graph.json","events_json":"https://pith.science/api/pith-number/R6APKS7T34X7NKFOYYN5A6CNVD/events.json","paper":"https://pith.science/paper/R6APKS7T"},"agent_actions":{"view_html":"https://pith.science/pith/R6APKS7T34X7NKFOYYN5A6CNVD","download_json":"https://pith.science/pith/R6APKS7T34X7NKFOYYN5A6CNVD.json","view_paper":"https://pith.science/paper/R6APKS7T","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2602.18037&json=true","fetch_graph":"https://pith.science/api/pith-number/R6APKS7T34X7NKFOYYN5A6CNVD/graph.json","fetch_events":"https://pith.science/api/pith-number/R6APKS7T34X7NKFOYYN5A6CNVD/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/R6APKS7T34X7NKFOYYN5A6CNVD/action/timestamp_anchor","attest_storage":"https://pith.science/pith/R6APKS7T34X7NKFOYYN5A6CNVD/action/storage_attestation","attest_author":"https://pith.science/pith/R6APKS7T34X7NKFOYYN5A6CNVD/action/author_attestation","sign_citation":"https://pith.science/pith/R6APKS7T34X7NKFOYYN5A6CNVD/action/citation_signature","submit_replication":"https://pith.science/pith/R6APKS7T34X7NKFOYYN5A6CNVD/action/replication_record"}},"created_at":"2026-07-07T00:15:52.603232+00:00","updated_at":"2026-07-07T00:15:52.603232+00:00"}