{"paper":{"title":"Anatomy of Unlearning: The Dual Impact of Fact Salience and Model Fine-Tuning","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","headline":"An SFT step on data to forget produces smoother unlearning and 10-50% higher retention than direct unlearning on pretrained models.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Alexander Panchenko, Andrey Savchenko, Anna Borisiuk, Elena Tutubalina","submitted_at":"2026-02-23T08:58:48Z","abstract_excerpt":"Machine Unlearning (MU) enables Large Language Models (LLMs) to remove unsafe or outdated information. However, existing work assumes that all facts are equally forgettable and largely ignores whether the forgotten knowledge originates from pretraining or supervised fine-tuning (SFT). In this paper, we introduce DUET (Dual Unlearning Evaluation across Training Stages), a benchmark of 28.6k Wikidata-derived triplets annotated with fact popularity using Wikipedia link counts and LLM-based salience scores. Our experiments show that pretrained and SFT models respond differently to unlearning. An S"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"An SFT step on the forget data yields smoother forgetting, more stable tuning, and 10-50% higher retention, while direct unlearning on pretrained models remains unstable and prone to relearning or catastrophic forgetting.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That differences in unlearning behavior are caused primarily by the presence or absence of an SFT stage rather than by other uncontrolled factors such as model scale, exact unlearning algorithm, or how salience scores correlate with actual memorization.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"SFT models forget facts more stably than pretrained models, with 10-50% higher retention of unrelated knowledge when using the DUET benchmark of 28.6k Wikidata triplets.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"An SFT step on data to forget produces smoother unlearning and 10-50% higher retention than direct unlearning on pretrained models.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"43ec0eb5729b60bcc9cb3f1968c596e40b2046c8ccb26761b60062b3045e46b6"},"source":{"id":"2602.19612","kind":"arxiv","version":5},"verdict":{"id":"a45662ce-b499-419e-80e9-fc852f5116c3","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T20:50:04.687150Z","strongest_claim":"An SFT step on the forget data yields smoother forgetting, more stable tuning, and 10-50% higher retention, while direct unlearning on pretrained models remains unstable and prone to relearning or catastrophic forgetting.","one_line_summary":"SFT models forget facts more stably than pretrained models, with 10-50% higher retention of unrelated knowledge when using the DUET benchmark of 28.6k Wikidata triplets.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That differences in unlearning behavior are caused primarily by the presence or absence of an SFT stage rather than by other uncontrolled factors such as model scale, exact unlearning algorithm, or how salience scores correlate with actual memorization.","pith_extraction_headline":"An SFT step on data to forget produces smoother unlearning and 10-50% higher retention than direct unlearning on pretrained models."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2602.19612/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"}