{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2024:R6LMVMYRQPXL7SCVEFHK4F7YWX","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":"dc7a30f5f9d624103362675113b078372199a9847637720efa402c24b63908d5","cross_cats_sorted":["cs.CL"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2024-11-10T11:07:34Z","title_canon_sha256":"1d38f9369c0c7d2737cedfcd051e6dad9873a5e5170503f1fdff580d6ef3110f"},"schema_version":"1.0","source":{"id":"2411.06424","kind":"arxiv","version":3}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2411.06424","created_at":"2026-07-05T11:17:37Z"},{"alias_kind":"arxiv_version","alias_value":"2411.06424v3","created_at":"2026-07-05T11:17:37Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2411.06424","created_at":"2026-07-05T11:17:37Z"},{"alias_kind":"pith_short_12","alias_value":"R6LMVMYRQPXL","created_at":"2026-07-05T11:17:37Z"},{"alias_kind":"pith_short_16","alias_value":"R6LMVMYRQPXL7SCV","created_at":"2026-07-05T11:17:37Z"},{"alias_kind":"pith_short_8","alias_value":"R6LMVMYR","created_at":"2026-07-05T11:17:37Z"}],"graph_snapshots":[{"event_id":"sha256:6720b5fe6b8ff386c7c5c476bf3bd832cf66630be938fd0cbd13b7c4d39af032","target":"graph","created_at":"2026-07-05T11:17:37Z","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":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"integrity":{"available":true,"clean":true,"detectors_run":[],"endpoint":"/pith/2411.06424/integrity.json","findings":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938","summary":{"advisory":0,"by_detector":{},"critical":0,"informational":0}},"paper":{"abstract_excerpt":"Safety fine-tuning algorithms reduce harmful outputs in language models, yet their mechanisms remain under-explored. Direct Preference Optimization (DPO) is a popular choice of algorithm, but prior explanations, attributing its effects solely to dampened toxic neurons in the MLP layers, are incomplete. In this study, we analyse four language models (Llama-3.1-8B, Gemma-2-2B, Mistral-7B, GPT-2-Medium) and show that toxic neurons only account for 2.5% to 24% of DPO's effects across models. Instead, DPO balances distributed activation shifts across all MLP neurons to create a net toxicity reducti","authors_text":"Adam Mahdi, Andrew Lee, Filip Sondej, Harry Mayne, Yushi Yang","cross_cats":["cs.CL"],"headline":"","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2024-11-10T11:07:34Z","title":"How Does DPO Reduce Toxicity? A Mechanistic Neuron-Level Analysis"},"references":{"count":0,"internal_anchors":0,"resolved_work":0,"sample":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2411.06424","kind":"arxiv","version":3},"verdict":{"created_at":null,"id":null,"model_set":{},"one_line_summary":"","pipeline_version":null,"pith_extraction_headline":"","strongest_claim":"","weakest_assumption":""}},"verdict_id":null}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:8092103e2d445e8b9db4ca6faa5a58987cb9d7c373f9e2fe76d6dbc193c8ceef","target":"record","created_at":"2026-07-05T11:17:37Z","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":"dc7a30f5f9d624103362675113b078372199a9847637720efa402c24b63908d5","cross_cats_sorted":["cs.CL"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2024-11-10T11:07:34Z","title_canon_sha256":"1d38f9369c0c7d2737cedfcd051e6dad9873a5e5170503f1fdff580d6ef3110f"},"schema_version":"1.0","source":{"id":"2411.06424","kind":"arxiv","version":3}},"canonical_sha256":"8f96cab31183eebfc855214eae17f8b5ead7563dcf0ea146003ed0020dd5a13d","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"8f96cab31183eebfc855214eae17f8b5ead7563dcf0ea146003ed0020dd5a13d","first_computed_at":"2026-07-05T11:17:37.345699Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-07-05T11:17:37.345699Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"sX5IfAb34n+Ohd2LFEyFlg8/bMw5saeeHsaNbGaqK1M6+lRrdfMUUvJp8hcsiBEOCoCdt484zFlBHWvsMNPmBA==","signature_status":"signed_v1","signed_at":"2026-07-05T11:17:37.346457Z","signed_message":"canonical_sha256_bytes"},"source_id":"2411.06424","source_kind":"arxiv","source_version":3}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:8092103e2d445e8b9db4ca6faa5a58987cb9d7c373f9e2fe76d6dbc193c8ceef","sha256:6720b5fe6b8ff386c7c5c476bf3bd832cf66630be938fd0cbd13b7c4d39af032"],"state_sha256":"e72b673a111ed33b19ff58a597a03cbc96e5f283fb5636a9904740a5dc49c0bb"}