{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:TJ5GE6UHFQ4JTWTHVCJ756K4IH","short_pith_number":"pith:TJ5GE6UH","schema_version":"1.0","canonical_sha256":"9a7a627a872c3899da67a893fef95c41ddafae7dd77b9fc902dc828379c69a5c","source":{"kind":"arxiv","id":"2406.16768","version":1},"attestation_state":"computed","paper":{"title":"WARP: On the Benefits of Weight Averaged Rewarded Policies","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Alexandre Ram\\'e, Arthur Douillard, Johan Ferret, L\\'eonard Hussenot, Nino Vieillard, Olivier Bachem, Pier Giuseppe Sessa, Pierre-Louis Cedoz, Robert Dadashi, Sertan Girgin","submitted_at":"2024-06-24T16:24:34Z","abstract_excerpt":"Reinforcement learning from human feedback (RLHF) aligns large language models (LLMs) by encouraging their generations to have high rewards, using a reward model trained on human preferences. To prevent the forgetting of pre-trained knowledge, RLHF usually incorporates a KL regularization; this forces the policy to remain close to its supervised fine-tuned initialization, though it hinders the reward optimization. To tackle the trade-off between KL and reward, in this paper we introduce a novel alignment strategy named Weight Averaged Rewarded Policies (WARP). WARP merges policies in the weigh"},"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":"2406.16768","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","primary_cat":"cs.LG","submitted_at":"2024-06-24T16:24:34Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"95564f0e799a83615219ebe3fa7bd3ad40d64670a757429340d55a86de3faf4f","abstract_canon_sha256":"03d227f539a117c89a23458d98d27687c7e9772413b081a3937512e293b7d1a8"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:36:03.352084Z","signature_b64":"fnDgwoA9TQJ2HkVdLnY6jgHHe3LCp5CjKULleBFAiKX8tYHHKVAx5Lzk/UlGssXI6Wf6YPAZ1Ak+fipNqWhdDw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"9a7a627a872c3899da67a893fef95c41ddafae7dd77b9fc902dc828379c69a5c","last_reissued_at":"2026-07-05T08:36:03.351599Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:36:03.351599Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"WARP: On the Benefits of Weight Averaged Rewarded Policies","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Alexandre Ram\\'e, Arthur Douillard, Johan Ferret, L\\'eonard Hussenot, Nino Vieillard, Olivier Bachem, Pier Giuseppe Sessa, Pierre-Louis Cedoz, Robert Dadashi, Sertan Girgin","submitted_at":"2024-06-24T16:24:34Z","abstract_excerpt":"Reinforcement learning from human feedback (RLHF) aligns large language models (LLMs) by encouraging their generations to have high rewards, using a reward model trained on human preferences. To prevent the forgetting of pre-trained knowledge, RLHF usually incorporates a KL regularization; this forces the policy to remain close to its supervised fine-tuned initialization, though it hinders the reward optimization. To tackle the trade-off between KL and reward, in this paper we introduce a novel alignment strategy named Weight Averaged Rewarded Policies (WARP). WARP merges policies in the weigh"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2406.16768","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/2406.16768/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":"2406.16768","created_at":"2026-07-05T08:36:03.351657+00:00"},{"alias_kind":"arxiv_version","alias_value":"2406.16768v1","created_at":"2026-07-05T08:36:03.351657+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2406.16768","created_at":"2026-07-05T08:36:03.351657+00:00"},{"alias_kind":"pith_short_12","alias_value":"TJ5GE6UHFQ4J","created_at":"2026-07-05T08:36:03.351657+00:00"},{"alias_kind":"pith_short_16","alias_value":"TJ5GE6UHFQ4JTWTH","created_at":"2026-07-05T08:36:03.351657+00:00"},{"alias_kind":"pith_short_8","alias_value":"TJ5GE6UH","created_at":"2026-07-05T08:36:03.351657+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":7,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.23995","citing_title":"EMAgnet: Parameter-Space EMA Regularization for Policy Gradient Self-Play in Large Games","ref_index":23,"is_internal_anchor":false},{"citing_arxiv_id":"2605.28751","citing_title":"Extrapolative Weight Averaging Reveals Correctness-Efficiency Frontiers in Code RL","ref_index":36,"is_internal_anchor":false},{"citing_arxiv_id":"2507.15698","citing_title":"CoLD: Counterfactually-Guided Length Debiasing for Process Reward Models in Mathematical Reasoning","ref_index":16,"is_internal_anchor":false},{"citing_arxiv_id":"2605.20408","citing_title":"Spectral Souping: A Unified Framework for Online Preference Alignment","ref_index":23,"is_internal_anchor":false},{"citing_arxiv_id":"2408.07666","citing_title":"Model Merging in LLMs, MLLMs, and Beyond: Methods, Theories, Applications and Opportunities","ref_index":177,"is_internal_anchor":false},{"citing_arxiv_id":"2605.00610","citing_title":"Decouple before Integration: Test-time Synthesis of SFT and RLVR Task Vectors","ref_index":51,"is_internal_anchor":false},{"citing_arxiv_id":"2604.13833","citing_title":"Robust Reward Modeling for Large Language Models via Causal Decomposition","ref_index":1,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/TJ5GE6UHFQ4JTWTHVCJ756K4IH","json":"https://pith.science/pith/TJ5GE6UHFQ4JTWTHVCJ756K4IH.json","graph_json":"https://pith.science/api/pith-number/TJ5GE6UHFQ4JTWTHVCJ756K4IH/graph.json","events_json":"https://pith.science/api/pith-number/TJ5GE6UHFQ4JTWTHVCJ756K4IH/events.json","paper":"https://pith.science/paper/TJ5GE6UH"},"agent_actions":{"view_html":"https://pith.science/pith/TJ5GE6UHFQ4JTWTHVCJ756K4IH","download_json":"https://pith.science/pith/TJ5GE6UHFQ4JTWTHVCJ756K4IH.json","view_paper":"https://pith.science/paper/TJ5GE6UH","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2406.16768&json=true","fetch_graph":"https://pith.science/api/pith-number/TJ5GE6UHFQ4JTWTHVCJ756K4IH/graph.json","fetch_events":"https://pith.science/api/pith-number/TJ5GE6UHFQ4JTWTHVCJ756K4IH/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/TJ5GE6UHFQ4JTWTHVCJ756K4IH/action/timestamp_anchor","attest_storage":"https://pith.science/pith/TJ5GE6UHFQ4JTWTHVCJ756K4IH/action/storage_attestation","attest_author":"https://pith.science/pith/TJ5GE6UHFQ4JTWTHVCJ756K4IH/action/author_attestation","sign_citation":"https://pith.science/pith/TJ5GE6UHFQ4JTWTHVCJ756K4IH/action/citation_signature","submit_replication":"https://pith.science/pith/TJ5GE6UHFQ4JTWTHVCJ756K4IH/action/replication_record"}},"created_at":"2026-07-05T08:36:03.351657+00:00","updated_at":"2026-07-05T08:36:03.351657+00:00"}