{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:KWFZ7LKCVDH2UJMDUDL7SIR6CL","short_pith_number":"pith:KWFZ7LKC","schema_version":"1.0","canonical_sha256":"558b9fad42a8cfaa2583a0d7f9223e12fd89bae3676dcccdca66ce398399cc8a","source":{"kind":"arxiv","id":"2605.16865","version":1},"attestation_state":"computed","paper":{"title":"MixSD: Mixed Contextual Self-Distillation for Knowledge Injection","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Aligning fine-tuning targets with a language model's own generation distribution prevents catastrophic forgetting of pretrained capabilities.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Jiarui Liu, Lechen Zhang, Mona Diab, Weihao Xuan, Yingheng Wang, Yinghui He, Yongjin Yang, Zhijing Jin","submitted_at":"2026-05-16T07:57:09Z","abstract_excerpt":"Supervised fine-tuning (SFT) is widely used to inject new knowledge into language models, but it often degrades pretrained capabilities such as reasoning and general-domain performance. We argue this forgetting arises because fine-tuning targets from humans or external systems diverge from the model's autoregressive distribution, forcing the optimizer to imitate low-probability token sequences. To address this problem, we propose MixSD, a simple external-teacher-free method for distribution-aligned knowledge injection. Instead of training on fixed targets, MixSD constructs supervision dynamica"},"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":true},"canonical_record":{"source":{"id":"2605.16865","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-05-16T07:57:09Z","cross_cats_sorted":[],"title_canon_sha256":"d91fa992df7169978a7e2800ca98bd851dbc95e62d7f4f7e1bb7ffd198ddadab","abstract_canon_sha256":"525996c2af35114b47f73523be53e16451c1d75fb47c025a9525adaa2e02f7bc"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-20T00:03:27.144573Z","signature_b64":"6CrSWbWEjtGUt/EDt1POJyxGjTbwszPnjmGHZRD8TuEl+XDIgrbq+9cCVcpMekOeh10lTvYxErGhM1C3pcUsAw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"558b9fad42a8cfaa2583a0d7f9223e12fd89bae3676dcccdca66ce398399cc8a","last_reissued_at":"2026-05-20T00:03:27.143636Z","signature_status":"signed_v1","first_computed_at":"2026-05-20T00:03:27.143636Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"MixSD: Mixed Contextual Self-Distillation for Knowledge Injection","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Aligning fine-tuning targets with a language model's own generation distribution prevents catastrophic forgetting of pretrained capabilities.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Jiarui Liu, Lechen Zhang, Mona Diab, Weihao Xuan, Yingheng Wang, Yinghui He, Yongjin Yang, Zhijing Jin","submitted_at":"2026-05-16T07:57:09Z","abstract_excerpt":"Supervised fine-tuning (SFT) is widely used to inject new knowledge into language models, but it often degrades pretrained capabilities such as reasoning and general-domain performance. We argue this forgetting arises because fine-tuning targets from humans or external systems diverge from the model's autoregressive distribution, forcing the optimizer to imitate low-probability token sequences. To address this problem, we propose MixSD, a simple external-teacher-free method for distribution-aligned knowledge injection. Instead of training on fixed targets, MixSD constructs supervision dynamica"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"aligning supervision with the model's native generation distribution is a simple and effective principle for knowledge injection that mitigates catastrophic forgetting.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The mixed supervision sequences preserve the factual learning signal while remaining substantially closer to the base model's distribution, as constructed from the expert and naive conditionals.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"MixSD achieves superior memorization-retention trade-off in knowledge injection by using mixed self-generated supervision from the base model's conditionals, retaining up to 100% held-out capability versus 1% for standard SFT.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Aligning fine-tuning targets with a language model's own generation distribution prevents catastrophic forgetting of pretrained capabilities.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"cd18781c7d4e522cdbaf1f2cb97332840812d3a736c9e59bbe319121bc17694c"},"source":{"id":"2605.16865","kind":"arxiv","version":1},"verdict":{"id":"7a497101-b14e-42e1-aa3c-7520a45d5071","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T21:05:54.150137Z","strongest_claim":"aligning supervision with the model's native generation distribution is a simple and effective principle for knowledge injection that mitigates catastrophic forgetting.","one_line_summary":"MixSD achieves superior memorization-retention trade-off in knowledge injection by using mixed self-generated supervision from the base model's conditionals, retaining up to 100% held-out capability versus 1% for standard SFT.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The mixed supervision sequences preserve the factual learning signal while remaining substantially closer to the base model's distribution, as constructed from the expert and naive conditionals.","pith_extraction_headline":"Aligning fine-tuning targets with a language model's own generation distribution prevents catastrophic forgetting of pretrained capabilities."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.16865/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_title_agreement","ran_at":"2026-05-19T21:31:19.200037Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T21:11:42.647492Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"claim_evidence","ran_at":"2026-05-19T18:41:56.302458Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T18:33:26.378248Z","status":"skipped","version":"1.0.0","findings_count":0}],"snapshot_sha256":"4dd803e5858d86c45e1f57b7be76b455ed8c54b4ac246289803a1793f97a7779"},"references":{"count":57,"sample":[{"doi":"","year":null,"title":"Measuring short-form factuality in large language models","work_id":"f8e490ab-7057-43fb-8c6d-06fc603836c7","ref_index":1,"cited_arxiv_id":"2411.04368","is_internal_anchor":true},{"doi":"","year":null,"title":"arXiv preprint arXiv:2502.20377 , year=","work_id":"253720f1-271f-4a18-b48e-080715a0305c","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"American Invitational Mathematics Examination (AIME) 2024 , author=","work_id":"ba78ac5d-070c-4d17-be42-6859a493bb86","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Measuring Mathematical Problem Solving With the MATH Dataset","work_id":"50652ac6-fb7c-4675-a2c2-159c241feb17","ref_index":4,"cited_arxiv_id":"2103.03874","is_internal_anchor":true},{"doi":"","year":null,"title":"Training Verifiers to Solve Math Word Problems","work_id":"acab1aa8-b4d6-40e0-a3ee-25341701dca2","ref_index":5,"cited_arxiv_id":"2110.14168","is_internal_anchor":true}],"resolved_work":57,"snapshot_sha256":"d2292109ecc0ab2337125fad49a0ad4c5ebcf954b5859f9e9ef9ef74b6499a04","internal_anchors":9},"formal_canon":{"evidence_count":1,"snapshot_sha256":"dfda7bacdc5f124f8b0b270e5cc87636fe885dbe6c185e2df4a58636e523ef42"},"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":"2605.16865","created_at":"2026-05-20T00:03:27.143797+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.16865v1","created_at":"2026-05-20T00:03:27.143797+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.16865","created_at":"2026-05-20T00:03:27.143797+00:00"},{"alias_kind":"pith_short_12","alias_value":"KWFZ7LKCVDH2","created_at":"2026-05-20T00:03:27.143797+00:00"},{"alias_kind":"pith_short_16","alias_value":"KWFZ7LKCVDH2UJMD","created_at":"2026-05-20T00:03:27.143797+00:00"},{"alias_kind":"pith_short_8","alias_value":"KWFZ7LKC","created_at":"2026-05-20T00:03:27.143797+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":1,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/KWFZ7LKCVDH2UJMDUDL7SIR6CL","json":"https://pith.science/pith/KWFZ7LKCVDH2UJMDUDL7SIR6CL.json","graph_json":"https://pith.science/api/pith-number/KWFZ7LKCVDH2UJMDUDL7SIR6CL/graph.json","events_json":"https://pith.science/api/pith-number/KWFZ7LKCVDH2UJMDUDL7SIR6CL/events.json","paper":"https://pith.science/paper/KWFZ7LKC"},"agent_actions":{"view_html":"https://pith.science/pith/KWFZ7LKCVDH2UJMDUDL7SIR6CL","download_json":"https://pith.science/pith/KWFZ7LKCVDH2UJMDUDL7SIR6CL.json","view_paper":"https://pith.science/paper/KWFZ7LKC","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.16865&json=true","fetch_graph":"https://pith.science/api/pith-number/KWFZ7LKCVDH2UJMDUDL7SIR6CL/graph.json","fetch_events":"https://pith.science/api/pith-number/KWFZ7LKCVDH2UJMDUDL7SIR6CL/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/KWFZ7LKCVDH2UJMDUDL7SIR6CL/action/timestamp_anchor","attest_storage":"https://pith.science/pith/KWFZ7LKCVDH2UJMDUDL7SIR6CL/action/storage_attestation","attest_author":"https://pith.science/pith/KWFZ7LKCVDH2UJMDUDL7SIR6CL/action/author_attestation","sign_citation":"https://pith.science/pith/KWFZ7LKCVDH2UJMDUDL7SIR6CL/action/citation_signature","submit_replication":"https://pith.science/pith/KWFZ7LKCVDH2UJMDUDL7SIR6CL/action/replication_record"}},"created_at":"2026-05-20T00:03:27.143797+00:00","updated_at":"2026-05-20T00:03:27.143797+00:00"}