{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2025:T7VQIBNLLTE2TBPWOMMR5JEKYR","short_pith_number":"pith:T7VQIBNL","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"},"canonical_sha256":"9feb0405ab5cc9a985f673191ea48ac45049de8d784429319118892e7c71d92e","source":{"kind":"arxiv","id":"2509.04259","version":1},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2509.04259","created_at":"2026-05-17T23:38:15Z"},{"alias_kind":"arxiv_version","alias_value":"2509.04259v1","created_at":"2026-05-17T23:38:15Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2509.04259","created_at":"2026-05-17T23:38:15Z"},{"alias_kind":"pith_short_12","alias_value":"T7VQIBNLLTE2","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"T7VQIBNLLTE2TBPW","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"T7VQIBNL","created_at":"2026-05-18T12:33:37Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2025:T7VQIBNLLTE2TBPWOMMR5JEKYR","target":"record","payload":{"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"},"canonical_sha256":"9feb0405ab5cc9a985f673191ea48ac45049de8d784429319118892e7c71d92e","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"},"source_kind":"arxiv","source_id":"2509.04259","source_version":1,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-17T23:38:15Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"rOYTpD7bUVUZ93/M0GM7xV2VTc8yT+yMpdfVTDajBm06ICb2tgwrEhgrSL86LSohbwlOrvb3QCGSoxIi4c9LAg==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-06-01T01:39:56.111525Z"},"content_sha256":"8d39c9a82013e9993fc0cb593a64dc5ce688d6ad071354f1f0ce2eab7fd53dc1","schema_version":"1.0","event_id":"sha256:8d39c9a82013e9993fc0cb593a64dc5ce688d6ad071354f1f0ce2eab7fd53dc1"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2025:T7VQIBNLLTE2TBPWOMMR5JEKYR","target":"graph","payload":{"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"},"verdict_id":"0ffaaabe-4c29-4279-a890-fe6cbee460ff"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-17T23:38:15Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"W/f2B6YxQDaKVxPn90OYIn473ShbbcKScLDUxE972T19xosWffNzZGBPdIikKQg38uMKGc/GnWZvDhLv+6axBQ==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-06-01T01:39:56.112721Z"},"content_sha256":"0c4ed9d8a03dd72c0b05d426ce666999eefffc92cdcd2e50adc68a6b1a8be5e8","schema_version":"1.0","event_id":"sha256:0c4ed9d8a03dd72c0b05d426ce666999eefffc92cdcd2e50adc68a6b1a8be5e8"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/T7VQIBNLLTE2TBPWOMMR5JEKYR/bundle.json","state_url":"https://pith.science/pith/T7VQIBNLLTE2TBPWOMMR5JEKYR/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/T7VQIBNLLTE2TBPWOMMR5JEKYR/bundle.json","status":"primary"}],"public_keys":[{"key_id":"pith-v1-2026-05","algorithm":"ed25519","format":"raw","public_key_b64":"stVStoiQhXFxp4s2pdzPNoqVNBMojDU/fJ2db5S3CbM=","public_key_hex":"b2d552b68890857171a78b36a5dccf368a953413288c353f7c9d9d6f94b709b3","fingerprint_sha256_b32_first128bits":"RVFV5Z2OI2J3ZUO7ERDEBCYNKS","fingerprint_sha256_hex":"8d4b5ee74e4693bcd1df2446408b0d54","rotates_at":null,"url":"https://pith.science/pith-signing-key.json","notes":"Pith uses this Ed25519 key to sign canonical record SHA-256 digests. Verify with: ed25519_verify(public_key, message=canonical_sha256_bytes, signature=base64decode(signature_b64))."}],"merge_version":"pith-open-graph-merge-v1","built_at":"2026-06-01T01:39:56Z","links":{"resolver":"https://pith.science/pith/T7VQIBNLLTE2TBPWOMMR5JEKYR","bundle":"https://pith.science/pith/T7VQIBNLLTE2TBPWOMMR5JEKYR/bundle.json","state":"https://pith.science/pith/T7VQIBNLLTE2TBPWOMMR5JEKYR/state.json","well_known_bundle":"https://pith.science/.well-known/pith/T7VQIBNLLTE2TBPWOMMR5JEKYR/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2025:T7VQIBNLLTE2TBPWOMMR5JEKYR","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"06de02bce6ff42333c41e0a972d2169fd6b955543aa8d13f5ef008acba3af663","cross_cats_sorted":[],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-09-04T14:38:08Z","title_canon_sha256":"9b5cf8274c7ae4584cabcb5cbf3d95444013f1dba694d01c57bdbdf3abbf1241"},"schema_version":"1.0","source":{"id":"2509.04259","kind":"arxiv","version":1}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2509.04259","created_at":"2026-05-17T23:38:15Z"},{"alias_kind":"arxiv_version","alias_value":"2509.04259v1","created_at":"2026-05-17T23:38:15Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2509.04259","created_at":"2026-05-17T23:38:15Z"},{"alias_kind":"pith_short_12","alias_value":"T7VQIBNLLTE2","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"T7VQIBNLLTE2TBPW","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"T7VQIBNL","created_at":"2026-05-18T12:33:37Z"}],"graph_snapshots":[{"event_id":"sha256:0c4ed9d8a03dd72c0b05d426ce666999eefffc92cdcd2e50adc68a6b1a8be5e8","target":"graph","created_at":"2026-05-17T23:38:15Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","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"},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"that the observed degree of forgetting is determined by the KL-divergence between fine-tuned and base policy evaluated on the new task"},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","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."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"On-policy RL forgets less than SFT because it selects the minimal-KL solution to new tasks among many possibilities."}],"snapshot_sha256":"1cd2cd214d3751b43abc9e702f65dc1c52df2b1ccb2867904c38d7ad1d4354c4"},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"paper":{"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","authors_text":"Idan Shenfeld, Jyothish Pari, Pulkit Agrawal","cross_cats":[],"headline":"On-policy RL forgets less than SFT because it selects the minimal-KL solution to new tasks among many possibilities.","license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-09-04T14:38:08Z","title":"RL's Razor: Why Online Reinforcement Learning Forgets Less"},"references":{"count":12,"internal_anchors":0,"resolved_work":12,"sample":[{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":1,"title":"Reinforcement fine-tuning naturally mitigates forgetting in continual post-training","work_id":"a9a5df4a-1aad-47a6-af40-807f0d9f3c5d","year":2019},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":2,"title":"We trained multiple models under a broad sweep of hyperparame- ters (see Table 2)","work_id":"eca2b784-272c-47e0-bc4f-48e7edfc12e7","year":null},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":3,"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","year":null},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":4,"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","year":2024},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":5,"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","year":null}],"snapshot_sha256":"6874af272df6a132b74ef5b3740c7407fc56d481f9563cd33b59e564bf7ad9f0"},"source":{"id":"2509.04259","kind":"arxiv","version":1},"verdict":{"created_at":"2026-05-17T04:58:40.180049Z","id":"0ffaaabe-4c29-4279-a890-fe6cbee460ff","model_set":{"reader":"grok-4.3"},"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","pith_extraction_headline":"On-policy RL forgets less than SFT because it selects the minimal-KL solution to new tasks among many possibilities.","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","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"}},"verdict_id":"0ffaaabe-4c29-4279-a890-fe6cbee460ff"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:8d39c9a82013e9993fc0cb593a64dc5ce688d6ad071354f1f0ce2eab7fd53dc1","target":"record","created_at":"2026-05-17T23:38:15Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"06de02bce6ff42333c41e0a972d2169fd6b955543aa8d13f5ef008acba3af663","cross_cats_sorted":[],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-09-04T14:38:08Z","title_canon_sha256":"9b5cf8274c7ae4584cabcb5cbf3d95444013f1dba694d01c57bdbdf3abbf1241"},"schema_version":"1.0","source":{"id":"2509.04259","kind":"arxiv","version":1}},"canonical_sha256":"9feb0405ab5cc9a985f673191ea48ac45049de8d784429319118892e7c71d92e","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"9feb0405ab5cc9a985f673191ea48ac45049de8d784429319118892e7c71d92e","first_computed_at":"2026-05-17T23:38:15.049025Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-17T23:38:15.049025Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"Enz8O/iGLfWK8MSD0xPsswqK3P6/tJ67w6oqPIyrDo20z2f3l1uKuBTtHoZ+uy2aJl096mtCpjmOaUixYAIJAg==","signature_status":"signed_v1","signed_at":"2026-05-17T23:38:15.049589Z","signed_message":"canonical_sha256_bytes"},"source_id":"2509.04259","source_kind":"arxiv","source_version":1}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:8d39c9a82013e9993fc0cb593a64dc5ce688d6ad071354f1f0ce2eab7fd53dc1","sha256:0c4ed9d8a03dd72c0b05d426ce666999eefffc92cdcd2e50adc68a6b1a8be5e8"],"state_sha256":"0e4d9b502f4ed814cf01c054934427d0688d6a08c4ea283825e4aaf5a35bf856"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"YzbrSB+vsyQtXG1FiC1Pq+8SoRbX31ZwuUsy+hA47EVKE442rFD1zxoCqaow6dEJ2IkbYxGBh6zowzonOsLxCQ==","signed_message":"bundle_sha256_bytes","signed_at":"2026-06-01T01:39:56.117574Z","bundle_sha256":"4ceb8705e1e45745df2285300edb7848df1df8bb6a88b243c716d2d91883b932"}}