{"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"}