{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:LC5NIJSY7Q7FSUOUYNFPE6MEPY","short_pith_number":"pith:LC5NIJSY","schema_version":"1.0","canonical_sha256":"58bad42658fc3e5951d4c34af279847e06c1700b3dc089e2f2516757497c46d5","source":{"kind":"arxiv","id":"2405.08448","version":1},"attestation_state":"computed","paper":{"title":"Understanding the performance gap between online and offline alignment algorithms","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Bernardo \\'Avila Pires, Daniele Calandriello, Daniel Zhaohan Guo, Eugene Tarassov, Michal Valko, R\\'emi Munos, Will Dabney, Yong Cheng, Yuan Cao, Yunhao Tang, Zeyu Zheng","submitted_at":"2024-05-14T09:12:30Z","abstract_excerpt":"Reinforcement learning from human feedback (RLHF) is the canonical framework for large language model alignment. However, rising popularity in offline alignment algorithms challenge the need for on-policy sampling in RLHF. Within the context of reward over-optimization, we start with an opening set of experiments that demonstrate the clear advantage of online methods over offline methods. This prompts us to investigate the causes to the performance discrepancy through a series of carefully designed experimental ablations. We show empirically that hypotheses such as offline data coverage and da"},"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":"2405.08448","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2024-05-14T09:12:30Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"c4cfb6a0633ea9911b3b27204728cab702a3e6850a8d8a4c99910b0b4ee43a0d","abstract_canon_sha256":"627c3366a71ab4a95e9295c2e039ebd9c4c2f4e068c4a52ef9a006d799e7ac0f"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:19:00.562614Z","signature_b64":"qDQ8L4WklT2jveY4yGetHbBAL21lN2VjeaGhQLjB96N09xcELRMDJrrp6UejyvjCd+vhBoWatraBiGathDPUDQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"58bad42658fc3e5951d4c34af279847e06c1700b3dc089e2f2516757497c46d5","last_reissued_at":"2026-07-05T08:19:00.562112Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:19:00.562112Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Understanding the performance gap between online and offline alignment algorithms","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Bernardo \\'Avila Pires, Daniele Calandriello, Daniel Zhaohan Guo, Eugene Tarassov, Michal Valko, R\\'emi Munos, Will Dabney, Yong Cheng, Yuan Cao, Yunhao Tang, Zeyu Zheng","submitted_at":"2024-05-14T09:12:30Z","abstract_excerpt":"Reinforcement learning from human feedback (RLHF) is the canonical framework for large language model alignment. However, rising popularity in offline alignment algorithms challenge the need for on-policy sampling in RLHF. Within the context of reward over-optimization, we start with an opening set of experiments that demonstrate the clear advantage of online methods over offline methods. This prompts us to investigate the causes to the performance discrepancy through a series of carefully designed experimental ablations. We show empirically that hypotheses such as offline data coverage and da"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2405.08448","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/2405.08448/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":"2405.08448","created_at":"2026-07-05T08:19:00.562172+00:00"},{"alias_kind":"arxiv_version","alias_value":"2405.08448v1","created_at":"2026-07-05T08:19:00.562172+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2405.08448","created_at":"2026-07-05T08:19:00.562172+00:00"},{"alias_kind":"pith_short_12","alias_value":"LC5NIJSY7Q7F","created_at":"2026-07-05T08:19:00.562172+00:00"},{"alias_kind":"pith_short_16","alias_value":"LC5NIJSY7Q7FSUOU","created_at":"2026-07-05T08:19:00.562172+00:00"},{"alias_kind":"pith_short_8","alias_value":"LC5NIJSY","created_at":"2026-07-05T08:19:00.562172+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":9,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.03489","citing_title":"Learn from Your Mistakes: Tree-like Self-Play for Secure Code LLMs","ref_index":15,"is_internal_anchor":false},{"citing_arxiv_id":"2505.17086","citing_title":"Advancing Multi-Agent RAG Systems with Minimalist Reinforcement Learning","ref_index":75,"is_internal_anchor":false},{"citing_arxiv_id":"2605.19033","citing_title":"RLFTSim: Realistic and Controllable Multi-Agent Traffic Simulation via Reinforcement Learning Fine-Tuning","ref_index":28,"is_internal_anchor":false},{"citing_arxiv_id":"2511.00066","citing_title":"Sharpness-Guided Group Relative Policy Optimization via Probability Shaping","ref_index":31,"is_internal_anchor":false},{"citing_arxiv_id":"2506.01939","citing_title":"Beyond the 80/20 Rule: High-Entropy Minority Tokens Drive Effective Reinforcement Learning for LLM Reasoning","ref_index":20,"is_internal_anchor":false},{"citing_arxiv_id":"2604.25872","citing_title":"When Errors Can Be Beneficial: A Categorization of Imperfect Rewards for Policy Gradient","ref_index":84,"is_internal_anchor":false},{"citing_arxiv_id":"2604.20051","citing_title":"Bootstrapping Post-training Signals for Open-ended Tasks via Rubric-based Self-play on Pre-training Text","ref_index":38,"is_internal_anchor":false},{"citing_arxiv_id":"2511.16624","citing_title":"SAM 3D: 3Dfy Anything in Images","ref_index":32,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06987","citing_title":"Response Time Enhances Alignment with Heterogeneous Preferences","ref_index":13,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/LC5NIJSY7Q7FSUOUYNFPE6MEPY","json":"https://pith.science/pith/LC5NIJSY7Q7FSUOUYNFPE6MEPY.json","graph_json":"https://pith.science/api/pith-number/LC5NIJSY7Q7FSUOUYNFPE6MEPY/graph.json","events_json":"https://pith.science/api/pith-number/LC5NIJSY7Q7FSUOUYNFPE6MEPY/events.json","paper":"https://pith.science/paper/LC5NIJSY"},"agent_actions":{"view_html":"https://pith.science/pith/LC5NIJSY7Q7FSUOUYNFPE6MEPY","download_json":"https://pith.science/pith/LC5NIJSY7Q7FSUOUYNFPE6MEPY.json","view_paper":"https://pith.science/paper/LC5NIJSY","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2405.08448&json=true","fetch_graph":"https://pith.science/api/pith-number/LC5NIJSY7Q7FSUOUYNFPE6MEPY/graph.json","fetch_events":"https://pith.science/api/pith-number/LC5NIJSY7Q7FSUOUYNFPE6MEPY/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/LC5NIJSY7Q7FSUOUYNFPE6MEPY/action/timestamp_anchor","attest_storage":"https://pith.science/pith/LC5NIJSY7Q7FSUOUYNFPE6MEPY/action/storage_attestation","attest_author":"https://pith.science/pith/LC5NIJSY7Q7FSUOUYNFPE6MEPY/action/author_attestation","sign_citation":"https://pith.science/pith/LC5NIJSY7Q7FSUOUYNFPE6MEPY/action/citation_signature","submit_replication":"https://pith.science/pith/LC5NIJSY7Q7FSUOUYNFPE6MEPY/action/replication_record"}},"created_at":"2026-07-05T08:19:00.562172+00:00","updated_at":"2026-07-05T08:19:00.562172+00:00"}