{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:OK2RXQX3XKQ6ZRI6M67WOX4EG6","short_pith_number":"pith:OK2RXQX3","schema_version":"1.0","canonical_sha256":"72b51bc2fbbaa1ecc51e67bf675f8437bf01f0b8eedc70abd7ee8dd515130274","source":{"kind":"arxiv","id":"2510.18874","version":3},"attestation_state":"computed","paper":{"title":"Retaining by Doing: The Role of On-Policy Data in Mitigating Forgetting","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Danqi Chen, Howard Chen, Karthik Narasimhan, Noam Razin","submitted_at":"2025-10-21T17:59:41Z","abstract_excerpt":"Adapting language models (LMs) to new tasks via post-training carries the risk of degrading existing capabilities -- a phenomenon classically known as catastrophic forgetting. In this paper, toward identifying guidelines for mitigating this phenomenon, we systematically compare the forgetting patterns of two widely adopted post-training methods: supervised fine-tuning (SFT) and reinforcement learning (RL). Our experiments reveal a consistent trend across LM families (Llama, Qwen) and tasks (instruction following, general knowledge, and arithmetic reasoning): RL leads to less forgetting than SF"},"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":"2510.18874","kind":"arxiv","version":3},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2025-10-21T17:59:41Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"406089db1e4544235f91fc7037237ae80de52a121f27e8395111d62d0a2f610e","abstract_canon_sha256":"90748e18964bf29367a1c0d6b9bf65cd4711903e1ddbe9b8dd1467df5f47be4e"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-29T01:14:25.142370Z","signature_b64":"izpt6DUTe0riw14TGgiYN5LUUnkWU9C66f+8UuE8yDPbWs1d/bapXdA7bn8Sdd7zIUJaIAOsJX9AjIX9MB1gBQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"72b51bc2fbbaa1ecc51e67bf675f8437bf01f0b8eedc70abd7ee8dd515130274","last_reissued_at":"2026-06-29T01:14:25.141813Z","signature_status":"signed_v1","first_computed_at":"2026-06-29T01:14:25.141813Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Retaining by Doing: The Role of On-Policy Data in Mitigating Forgetting","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Danqi Chen, Howard Chen, Karthik Narasimhan, Noam Razin","submitted_at":"2025-10-21T17:59:41Z","abstract_excerpt":"Adapting language models (LMs) to new tasks via post-training carries the risk of degrading existing capabilities -- a phenomenon classically known as catastrophic forgetting. In this paper, toward identifying guidelines for mitigating this phenomenon, we systematically compare the forgetting patterns of two widely adopted post-training methods: supervised fine-tuning (SFT) and reinforcement learning (RL). Our experiments reveal a consistent trend across LM families (Llama, Qwen) and tasks (instruction following, general knowledge, and arithmetic reasoning): RL leads to less forgetting than SF"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2510.18874","kind":"arxiv","version":3},"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/2510.18874/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":"2510.18874","created_at":"2026-06-29T01:14:25.141875+00:00"},{"alias_kind":"arxiv_version","alias_value":"2510.18874v3","created_at":"2026-06-29T01:14:25.141875+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2510.18874","created_at":"2026-06-29T01:14:25.141875+00:00"},{"alias_kind":"pith_short_12","alias_value":"OK2RXQX3XKQ6","created_at":"2026-06-29T01:14:25.141875+00:00"},{"alias_kind":"pith_short_16","alias_value":"OK2RXQX3XKQ6ZRI6","created_at":"2026-06-29T01:14:25.141875+00:00"},{"alias_kind":"pith_short_8","alias_value":"OK2RXQX3","created_at":"2026-06-29T01:14:25.141875+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":14,"internal_anchor_count":14,"sample":[{"citing_arxiv_id":"2606.30445","citing_title":"When Does Online Imitation Learning Help in LLM Post-Training? The Role of (Non-)Realizability Beyond Horizon","ref_index":5,"is_internal_anchor":true},{"citing_arxiv_id":"2605.22675","citing_title":"Self-Policy Distillation via Capability-Selective Subspace Projection","ref_index":15,"is_internal_anchor":true},{"citing_arxiv_id":"2602.10503","citing_title":"Towards Long-Lived Robots: Continual Learning VLA Models via Reinforcement Fine-Tuning","ref_index":10,"is_internal_anchor":true},{"citing_arxiv_id":"2605.19433","citing_title":"Backtracking When It Strays: Mitigating Dual Exposure Biases in LLM Reasoning Distillation","ref_index":4,"is_internal_anchor":true},{"citing_arxiv_id":"2601.14249","citing_title":"Which Reasoning Trajectories Teach Students to Reason Better? A Simple Metric of Informative Alignment","ref_index":5,"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":6,"is_internal_anchor":true},{"citing_arxiv_id":"2605.09725","citing_title":"On-Policy Distillation with Best-of-N Teacher Rollout Selection","ref_index":4,"is_internal_anchor":true},{"citing_arxiv_id":"2605.13165","citing_title":"STOP: Structured On-Policy Pruning of Long-Form Reasoning in Low-Data Regimes","ref_index":52,"is_internal_anchor":true},{"citing_arxiv_id":"2605.09725","citing_title":"On-Policy Distillation with Best-of-N Teacher Rollout Selection","ref_index":4,"is_internal_anchor":true},{"citing_arxiv_id":"2605.05732","citing_title":"CRAFT: Forgetting-Aware Intervention-Based Adaptation for Continual Learning","ref_index":4,"is_internal_anchor":true},{"citing_arxiv_id":"2605.04468","citing_title":"Stabilizing LLM Supervised Fine-Tuning via Explicit Distributional Control","ref_index":3,"is_internal_anchor":true},{"citing_arxiv_id":"2605.05732","citing_title":"CRAFT: Forgetting-Aware Intervention-Based Adaptation for Continual Learning","ref_index":4,"is_internal_anchor":true},{"citing_arxiv_id":"2604.05117","citing_title":"Watch Before You Answer: Learning from Visually Grounded Post-Training","ref_index":13,"is_internal_anchor":true},{"citing_arxiv_id":"2604.14518","citing_title":"Mind DeepResearch Technical Report","ref_index":1,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/OK2RXQX3XKQ6ZRI6M67WOX4EG6","json":"https://pith.science/pith/OK2RXQX3XKQ6ZRI6M67WOX4EG6.json","graph_json":"https://pith.science/api/pith-number/OK2RXQX3XKQ6ZRI6M67WOX4EG6/graph.json","events_json":"https://pith.science/api/pith-number/OK2RXQX3XKQ6ZRI6M67WOX4EG6/events.json","paper":"https://pith.science/paper/OK2RXQX3"},"agent_actions":{"view_html":"https://pith.science/pith/OK2RXQX3XKQ6ZRI6M67WOX4EG6","download_json":"https://pith.science/pith/OK2RXQX3XKQ6ZRI6M67WOX4EG6.json","view_paper":"https://pith.science/paper/OK2RXQX3","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2510.18874&json=true","fetch_graph":"https://pith.science/api/pith-number/OK2RXQX3XKQ6ZRI6M67WOX4EG6/graph.json","fetch_events":"https://pith.science/api/pith-number/OK2RXQX3XKQ6ZRI6M67WOX4EG6/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/OK2RXQX3XKQ6ZRI6M67WOX4EG6/action/timestamp_anchor","attest_storage":"https://pith.science/pith/OK2RXQX3XKQ6ZRI6M67WOX4EG6/action/storage_attestation","attest_author":"https://pith.science/pith/OK2RXQX3XKQ6ZRI6M67WOX4EG6/action/author_attestation","sign_citation":"https://pith.science/pith/OK2RXQX3XKQ6ZRI6M67WOX4EG6/action/citation_signature","submit_replication":"https://pith.science/pith/OK2RXQX3XKQ6ZRI6M67WOX4EG6/action/replication_record"}},"created_at":"2026-06-29T01:14:25.141875+00:00","updated_at":"2026-06-29T01:14:25.141875+00:00"}