{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2024:OUY3FJZZ64YXYLASJSH3UPIBMZ","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":"33132e33fc41748dd9cbaae0d98204eb9a4368ecb3197d78ea82a656e5eaa285","cross_cats_sorted":["cs.AI","cs.CL","stat.ML"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2024-04-08T21:05:42Z","title_canon_sha256":"3b37a71945c268e2b22937067c4ebb45934b5b6fca9614682c71e35e450d1684"},"schema_version":"1.0","source":{"id":"2404.05868","kind":"arxiv","version":2}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2404.05868","created_at":"2026-05-17T23:38:46Z"},{"alias_kind":"arxiv_version","alias_value":"2404.05868v2","created_at":"2026-05-17T23:38:46Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2404.05868","created_at":"2026-05-17T23:38:46Z"},{"alias_kind":"pith_short_12","alias_value":"OUY3FJZZ64YX","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"OUY3FJZZ64YXYLAS","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"OUY3FJZZ","created_at":"2026-05-18T12:33:37Z"}],"graph_snapshots":[{"event_id":"sha256:57449f81bbefcf453cdb51ff54fb47777307aafa14417f1c2cfbf9438f998c15","target":"graph","created_at":"2026-05-17T23:38:46Z","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":"Remarkably, on TOFU, NPO-based methods are the first to achieve reasonable unlearning results in forgetting 50% (or more) of the training data, whereas existing methods already struggle with forgetting 10% of training data."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"The assumption that results on the synthetic data and TOFU benchmark will generalize to real-world unlearning of sensitive data in large production LLMs without introducing new failure modes."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"NPO enables stable unlearning of 50%+ training data in LLMs on TOFU by making collapse exponentially slower than gradient ascent, preserving sensible outputs where prior methods fail."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Negative Preference Optimization unlearns large portions of LLM training data without catastrophic collapse."}],"snapshot_sha256":"6bb5a326d3b80ea65e0cabcdf28a1990f28c4262620658dc3808c7b37085c68c"},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"paper":{"abstract_excerpt":"Large Language Models (LLMs) often memorize sensitive, private, or copyrighted data during pre-training. LLM unlearning aims to eliminate the influence of undesirable data from the pre-trained model while preserving the model's utilities on other tasks. Several practical methods have recently been proposed for LLM unlearning, mostly based on gradient ascent (GA) on the loss of undesirable data. However, on certain unlearning tasks, these methods either fail to effectively unlearn the target data or suffer from catastrophic collapse -- a drastic degradation of the model's utilities.\n  In this p","authors_text":"Licong Lin, Ruiqi Zhang, Song Mei, Yu Bai","cross_cats":["cs.AI","cs.CL","stat.ML"],"headline":"Negative Preference Optimization unlearns large portions of LLM training data without catastrophic collapse.","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2024-04-08T21:05:42Z","title":"Negative Preference Optimization: From Catastrophic Collapse to Effective Unlearning"},"references":{"count":34,"internal_anchors":7,"resolved_work":34,"sample":[{"cited_arxiv_id":"2204.05862","doi":"","is_internal_anchor":true,"ref_index":1,"title":"Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback","work_id":"a1f2574b-a899-4713-be60-c87ba332656c","year":null},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":2,"title":"Machine unlearning","work_id":"ec065f1e-75e5-43a4-a873-1a134cd387db","year":2021},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":3,"title":"Towards making systems forget with machine unlearning","work_id":"f94eb5f4-9bee-4d94-a99f-86164ad870fc","year":2015},{"cited_arxiv_id":"2202.07646","doi":"","is_internal_anchor":true,"ref_index":4,"title":"Quantifying Memorization Across Neural Language Models","work_id":"35487ec1-b90b-4ace-95bd-1bce30064b2e","year":null},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":5,"title":"Unlearn what you want to forget: Efficient unlearning for llms","work_id":"903757fe-c1e7-41c0-9760-28e13aec8628","year":null}],"snapshot_sha256":"bdf084fcd3e8d0b818522ea3136779ef94393f926f2c06111194d39a27bfb3d4"},"source":{"id":"2404.05868","kind":"arxiv","version":2},"verdict":{"created_at":"2026-05-16T22:23:06.857645Z","id":"c2db78a8-25e3-4edc-8ea2-860620f78f4a","model_set":{"reader":"grok-4.3"},"one_line_summary":"NPO enables stable unlearning of 50%+ training data in LLMs on TOFU by making collapse exponentially slower than gradient ascent, preserving sensible outputs where prior methods fail.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"Negative Preference Optimization unlearns large portions of LLM training data without catastrophic collapse.","strongest_claim":"Remarkably, on TOFU, NPO-based methods are the first to achieve reasonable unlearning results in forgetting 50% (or more) of the training data, whereas existing methods already struggle with forgetting 10% of training data.","weakest_assumption":"The assumption that results on the synthetic data and TOFU benchmark will generalize to real-world unlearning of sensitive data in large production LLMs without introducing new failure modes."}},"verdict_id":"c2db78a8-25e3-4edc-8ea2-860620f78f4a"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:bc8d5128504b6c8361fa88197d56b17ff2dae5a55e699c01af43cbd5d5ebe961","target":"record","created_at":"2026-05-17T23:38:46Z","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":"33132e33fc41748dd9cbaae0d98204eb9a4368ecb3197d78ea82a656e5eaa285","cross_cats_sorted":["cs.AI","cs.CL","stat.ML"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2024-04-08T21:05:42Z","title_canon_sha256":"3b37a71945c268e2b22937067c4ebb45934b5b6fca9614682c71e35e450d1684"},"schema_version":"1.0","source":{"id":"2404.05868","kind":"arxiv","version":2}},"canonical_sha256":"7531b2a739f7317c2c124c8fba3d01664e0a38bb3b28428c4c26f40666a89ed5","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"7531b2a739f7317c2c124c8fba3d01664e0a38bb3b28428c4c26f40666a89ed5","first_computed_at":"2026-05-17T23:38:46.388851Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-17T23:38:46.388851Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"7lEYEJL6O4pXtbfGQvZnUCxFmB1i9yIKKfIkq/6zpFAVrv38SiXFkx34oxW3w0wsiGf3CkT8SdNogYeSpIaQBw==","signature_status":"signed_v1","signed_at":"2026-05-17T23:38:46.389295Z","signed_message":"canonical_sha256_bytes"},"source_id":"2404.05868","source_kind":"arxiv","source_version":2}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:bc8d5128504b6c8361fa88197d56b17ff2dae5a55e699c01af43cbd5d5ebe961","sha256:57449f81bbefcf453cdb51ff54fb47777307aafa14417f1c2cfbf9438f998c15"],"state_sha256":"cb68b580169126a16a28da6b55b96e2e316dd4ca1f0f2a165a9f6eef0a4eddea"}