{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:T7VQIBNLLTE2TBPWOMMR5JEKYR","short_pith_number":"pith:T7VQIBNL","schema_version":"1.0","canonical_sha256":"9feb0405ab5cc9a985f673191ea48ac45049de8d784429319118892e7c71d92e","source":{"kind":"arxiv","id":"2509.04259","version":1},"attestation_state":"computed","paper":{"title":"RL's Razor: Why Online Reinforcement Learning Forgets Less","license":"http://creativecommons.org/licenses/by/4.0/","headline":"On-policy RL forgets less than SFT because it selects the minimal-KL solution to new tasks among many possibilities.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Idan Shenfeld, Jyothish Pari, Pulkit Agrawal","submitted_at":"2025-09-04T14:38:08Z","abstract_excerpt":"Comparison of fine-tuning models with reinforcement learning (RL) and supervised fine-tuning (SFT) reveals that, despite similar performance at a new task, RL preserves prior knowledge and capabilities significantly better. We find that the degree of forgetting is determined by the distributional shift, measured as the KL-divergence between the fine-tuned and base policy evaluated on the new task. Our analysis reveals that on-policy RL is implicitly biased towards KL-minimal solutions among the many that solve the new task, whereas SFT can converge to distributions arbitrarily far from the bas"},"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":true,"formal_links_present":false},"canonical_record":{"source":{"id":"2509.04259","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-09-04T14:38:08Z","cross_cats_sorted":[],"title_canon_sha256":"9b5cf8274c7ae4584cabcb5cbf3d95444013f1dba694d01c57bdbdf3abbf1241","abstract_canon_sha256":"06de02bce6ff42333c41e0a972d2169fd6b955543aa8d13f5ef008acba3af663"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:38:15.049589Z","signature_b64":"Enz8O/iGLfWK8MSD0xPsswqK3P6/tJ67w6oqPIyrDo20z2f3l1uKuBTtHoZ+uy2aJl096mtCpjmOaUixYAIJAg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"9feb0405ab5cc9a985f673191ea48ac45049de8d784429319118892e7c71d92e","last_reissued_at":"2026-05-17T23:38:15.049025Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:38:15.049025Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"RL's Razor: Why Online Reinforcement Learning Forgets Less","license":"http://creativecommons.org/licenses/by/4.0/","headline":"On-policy RL forgets less than SFT because it selects the minimal-KL solution to new tasks among many possibilities.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Idan Shenfeld, Jyothish Pari, Pulkit Agrawal","submitted_at":"2025-09-04T14:38:08Z","abstract_excerpt":"Comparison of fine-tuning models with reinforcement learning (RL) and supervised fine-tuning (SFT) reveals that, despite similar performance at a new task, RL preserves prior knowledge and capabilities significantly better. We find that the degree of forgetting is determined by the distributional shift, measured as the KL-divergence between the fine-tuned and base policy evaluated on the new task. Our analysis reveals that on-policy RL is implicitly biased towards KL-minimal solutions among the many that solve the new task, whereas SFT can converge to distributions arbitrarily far from the bas"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"on-policy RL is implicitly biased towards KL-minimal solutions among the many that solve the new task, whereas SFT can converge to distributions arbitrarily far from the base model","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"that the observed degree of forgetting is determined by the KL-divergence between fine-tuned and base policy evaluated on the new task","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Online RL fine-tuning forgets less than SFT because it is implicitly biased toward KL-minimal solutions among all policies that solve the new task.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"On-policy RL forgets less than SFT because it selects the minimal-KL solution to new tasks among many possibilities.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"1cd2cd214d3751b43abc9e702f65dc1c52df2b1ccb2867904c38d7ad1d4354c4"},"source":{"id":"2509.04259","kind":"arxiv","version":1},"verdict":{"id":"0ffaaabe-4c29-4279-a890-fe6cbee460ff","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-17T04:58:40.180049Z","strongest_claim":"on-policy RL is implicitly biased towards KL-minimal solutions among the many that solve the new task, whereas SFT can converge to distributions arbitrarily far from the base model","one_line_summary":"Online RL fine-tuning forgets less than SFT because it is implicitly biased toward KL-minimal solutions among all policies that solve the new task.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"that the observed degree of forgetting is determined by the KL-divergence between fine-tuned and base policy evaluated on the new task","pith_extraction_headline":"On-policy RL forgets less than SFT because it selects the minimal-KL solution to new tasks among many possibilities."},"references":{"count":12,"sample":[{"doi":"","year":2019,"title":"Reinforcement fine-tuning naturally mitigates forgetting in continual post-training","work_id":"a9a5df4a-1aad-47a6-af40-807f0d9f3c5d","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"We trained multiple models under a broad sweep of hyperparame- ters (see Table 2)","work_id":"eca2b784-272c-47e0-bc4f-48e7edfc12e7","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"For Math and Science Q&A, accuracy was measured by comparing the model’s final answer to the ground truth, ignoring intermediate reasoning chains","work_id":"3a5d6055-0c14-467c-a966-4ab2d1964c5e","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"We assessed performance on unrelated benchmarks as described in Section 3.1, using the Language Model Evaluation Harness (Gao et al., 2024)","work_id":"150ae688-e43a-4ab3-9ba4-ab75604f40e4","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"From the trained models, we retained only those lying within 2 accuracy points of the Pareto frontier","work_id":"dc832f3b-29ae-453d-b874-a1e23e5e0e06","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":12,"snapshot_sha256":"6874af272df6a132b74ef5b3740c7407fc56d481f9563cd33b59e564bf7ad9f0","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":"2509.04259","created_at":"2026-05-17T23:38:15.049124+00:00"},{"alias_kind":"arxiv_version","alias_value":"2509.04259v1","created_at":"2026-05-17T23:38:15.049124+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2509.04259","created_at":"2026-05-17T23:38:15.049124+00:00"},{"alias_kind":"pith_short_12","alias_value":"T7VQIBNLLTE2","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"T7VQIBNLLTE2TBPW","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"T7VQIBNL","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":26,"internal_anchor_count":26,"sample":[{"citing_arxiv_id":"2602.10503","citing_title":"Towards Long-Lived Robots: Continual Learning VLA Models via Reinforcement Fine-Tuning","ref_index":61,"is_internal_anchor":true},{"citing_arxiv_id":"2602.03839","citing_title":"Understanding and Exploiting Weight Update Sparsity for Communication-Efficient Distributed RL","ref_index":30,"is_internal_anchor":true},{"citing_arxiv_id":"2605.08879","citing_title":"Preserving Foundational Capabilities in Flow-Matching VLAs through Conservative SFT","ref_index":8,"is_internal_anchor":true},{"citing_arxiv_id":"2605.18903","citing_title":"Reasoning Portability: Guiding Continual Learning for MLLMs in the RLVR Era","ref_index":42,"is_internal_anchor":true},{"citing_arxiv_id":"2509.23629","citing_title":"Emergent Slow Thinking in LLMs as Inverse Tree Freezing","ref_index":34,"is_internal_anchor":true},{"citing_arxiv_id":"2512.11470","citing_title":"Rethinking Expert Trajectory Utilization in LLM Post-training for Mathematical Reasoning","ref_index":35,"is_internal_anchor":true},{"citing_arxiv_id":"2605.12484","citing_title":"Learning, Fast and Slow: Towards LLMs That Adapt Continually","ref_index":53,"is_internal_anchor":true},{"citing_arxiv_id":"2604.14164","citing_title":"How to Fine-Tune a Reasoning Model? A Teacher-Student Cooperation Framework to Synthesize Student-Consistent SFT Data","ref_index":42,"is_internal_anchor":true},{"citing_arxiv_id":"2605.09725","citing_title":"On-Policy Distillation with Best-of-N Teacher Rollout Selection","ref_index":37,"is_internal_anchor":true},{"citing_arxiv_id":"2605.13080","citing_title":"Learning to See What You Need: Gaze Attention for Multimodal Large Language Models","ref_index":105,"is_internal_anchor":true},{"citing_arxiv_id":"2604.03179","citing_title":"Understanding the Role of Hallucination in Reinforcement Post-Training of Multimodal Reasoning Models","ref_index":26,"is_internal_anchor":true},{"citing_arxiv_id":"2605.10973","citing_title":"Rotation-Preserving Supervised Fine-Tuning","ref_index":35,"is_internal_anchor":true},{"citing_arxiv_id":"2605.12484","citing_title":"Learning, Fast and Slow: Towards LLMs That Adapt Continually","ref_index":52,"is_internal_anchor":true},{"citing_arxiv_id":"2605.08879","citing_title":"Preserving Foundational Capabilities in Flow-Matching VLAs through Conservative SFT","ref_index":8,"is_internal_anchor":true},{"citing_arxiv_id":"2605.09725","citing_title":"On-Policy Distillation with Best-of-N Teacher Rollout Selection","ref_index":37,"is_internal_anchor":true},{"citing_arxiv_id":"2605.10158","citing_title":"Unsupervised Process Reward Models","ref_index":61,"is_internal_anchor":true},{"citing_arxiv_id":"2604.23121","citing_title":"Breaking Lock-In: Preserving Steerability under Low-Data VLA Post-Training","ref_index":17,"is_internal_anchor":true},{"citing_arxiv_id":"2605.06523","citing_title":"On the Implicit Reward Overfitting and the Low-rank Dynamics in RLVR","ref_index":31,"is_internal_anchor":true},{"citing_arxiv_id":"2605.05732","citing_title":"CRAFT: Forgetting-Aware Intervention-Based Adaptation for Continual Learning","ref_index":14,"is_internal_anchor":true},{"citing_arxiv_id":"2605.04468","citing_title":"Stabilizing LLM Supervised Fine-Tuning via Explicit Distributional Control","ref_index":1,"is_internal_anchor":true},{"citing_arxiv_id":"2604.13010","citing_title":"Lightning OPD: Efficient Post-Training for Large Reasoning Models with Offline On-Policy Distillation","ref_index":20,"is_internal_anchor":true},{"citing_arxiv_id":"2605.05732","citing_title":"CRAFT: Forgetting-Aware Intervention-Based Adaptation for Continual Learning","ref_index":14,"is_internal_anchor":true},{"citing_arxiv_id":"2604.13010","citing_title":"Lightning OPD: Efficient Post-Training for Large Reasoning Models with Offline On-Policy Distillation","ref_index":20,"is_internal_anchor":true},{"citing_arxiv_id":"2604.20244","citing_title":"Hybrid Policy Distillation for LLMs","ref_index":12,"is_internal_anchor":true},{"citing_arxiv_id":"2605.02572","citing_title":"On Training Large Language Models for Long-Horizon Tasks: An Empirical Study of Horizon Length","ref_index":83,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/T7VQIBNLLTE2TBPWOMMR5JEKYR","json":"https://pith.science/pith/T7VQIBNLLTE2TBPWOMMR5JEKYR.json","graph_json":"https://pith.science/api/pith-number/T7VQIBNLLTE2TBPWOMMR5JEKYR/graph.json","events_json":"https://pith.science/api/pith-number/T7VQIBNLLTE2TBPWOMMR5JEKYR/events.json","paper":"https://pith.science/paper/T7VQIBNL"},"agent_actions":{"view_html":"https://pith.science/pith/T7VQIBNLLTE2TBPWOMMR5JEKYR","download_json":"https://pith.science/pith/T7VQIBNLLTE2TBPWOMMR5JEKYR.json","view_paper":"https://pith.science/paper/T7VQIBNL","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2509.04259&json=true","fetch_graph":"https://pith.science/api/pith-number/T7VQIBNLLTE2TBPWOMMR5JEKYR/graph.json","fetch_events":"https://pith.science/api/pith-number/T7VQIBNLLTE2TBPWOMMR5JEKYR/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/T7VQIBNLLTE2TBPWOMMR5JEKYR/action/timestamp_anchor","attest_storage":"https://pith.science/pith/T7VQIBNLLTE2TBPWOMMR5JEKYR/action/storage_attestation","attest_author":"https://pith.science/pith/T7VQIBNLLTE2TBPWOMMR5JEKYR/action/author_attestation","sign_citation":"https://pith.science/pith/T7VQIBNLLTE2TBPWOMMR5JEKYR/action/citation_signature","submit_replication":"https://pith.science/pith/T7VQIBNLLTE2TBPWOMMR5JEKYR/action/replication_record"}},"created_at":"2026-05-17T23:38:15.049124+00:00","updated_at":"2026-05-17T23:38:15.049124+00:00"}