{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:JVM6ZPKQ4OZUBK2NKACPDRHMCQ","short_pith_number":"pith:JVM6ZPKQ","schema_version":"1.0","canonical_sha256":"4d59ecbd50e3b340ab4d5004f1c4ec143621f701bd7bb4c2a18fa30532879c72","source":{"kind":"arxiv","id":"2607.18639","version":1},"attestation_state":"computed","paper":{"title":"Mark, Don't Erase: Token Inoculation for Dual-Use Knowledge in LLMs","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Dongyoon Han, Sangdoo Yun, Seunghyun Lee","submitted_at":"2026-07-21T02:15:39Z","abstract_excerpt":"Safety interventions on dual-use knowledge typically choose between destroying hazardous content (e.g., unlearning, filtering) and suppressing it at the output layer (e.g., refusal training); both pay a tax in adjacent-domain competence or over-refusal. We argue that the right operation is conditioning, not reduction: we show that hazardous knowledge can be retained in the model and behaviorally gated by a privileged control token. Our method, Token Inoculation, introduces a binding-and-branching approach. First, during continued pre-training, we mark hazardous content by inserting a special t"},"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":"2607.18639","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-07-21T02:15:39Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"f536f49fb78eae99dc1c3e1419d9de3e251b2bf33881f0c99b1dba06652b5d49","abstract_canon_sha256":"52ad76c4ac1b12ebcf44d6341531508a2db9c3c4cada4642682cd64d11a9b384"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-22T00:22:58.063783Z","signature_b64":"ThHIL6qYs7h49deVqp0u4QDq5Y912qb+ALTMF0c3qZx/q7h+lzbr/YRHgILnVQUM+xT+/baX/7Ga/eEb/6oLDQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"4d59ecbd50e3b340ab4d5004f1c4ec143621f701bd7bb4c2a18fa30532879c72","last_reissued_at":"2026-07-22T00:22:58.062951Z","signature_status":"signed_v1","first_computed_at":"2026-07-22T00:22:58.062951Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Mark, Don't Erase: Token Inoculation for Dual-Use Knowledge in LLMs","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Dongyoon Han, Sangdoo Yun, Seunghyun Lee","submitted_at":"2026-07-21T02:15:39Z","abstract_excerpt":"Safety interventions on dual-use knowledge typically choose between destroying hazardous content (e.g., unlearning, filtering) and suppressing it at the output layer (e.g., refusal training); both pay a tax in adjacent-domain competence or over-refusal. We argue that the right operation is conditioning, not reduction: we show that hazardous knowledge can be retained in the model and behaviorally gated by a privileged control token. Our method, Token Inoculation, introduces a binding-and-branching approach. First, during continued pre-training, we mark hazardous content by inserting a special t"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2607.18639","kind":"arxiv","version":1},"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/2607.18639/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":"2607.18639","created_at":"2026-07-22T00:22:58.063386+00:00"},{"alias_kind":"arxiv_version","alias_value":"2607.18639v1","created_at":"2026-07-22T00:22:58.063386+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2607.18639","created_at":"2026-07-22T00:22:58.063386+00:00"},{"alias_kind":"pith_short_12","alias_value":"JVM6ZPKQ4OZU","created_at":"2026-07-22T00:22:58.063386+00:00"},{"alias_kind":"pith_short_16","alias_value":"JVM6ZPKQ4OZUBK2N","created_at":"2026-07-22T00:22:58.063386+00:00"},{"alias_kind":"pith_short_8","alias_value":"JVM6ZPKQ","created_at":"2026-07-22T00:22:58.063386+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/JVM6ZPKQ4OZUBK2NKACPDRHMCQ","json":"https://pith.science/pith/JVM6ZPKQ4OZUBK2NKACPDRHMCQ.json","graph_json":"https://pith.science/api/pith-number/JVM6ZPKQ4OZUBK2NKACPDRHMCQ/graph.json","events_json":"https://pith.science/api/pith-number/JVM6ZPKQ4OZUBK2NKACPDRHMCQ/events.json","paper":"https://pith.science/paper/JVM6ZPKQ"},"agent_actions":{"view_html":"https://pith.science/pith/JVM6ZPKQ4OZUBK2NKACPDRHMCQ","download_json":"https://pith.science/pith/JVM6ZPKQ4OZUBK2NKACPDRHMCQ.json","view_paper":"https://pith.science/paper/JVM6ZPKQ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2607.18639&json=true","fetch_graph":"https://pith.science/api/pith-number/JVM6ZPKQ4OZUBK2NKACPDRHMCQ/graph.json","fetch_events":"https://pith.science/api/pith-number/JVM6ZPKQ4OZUBK2NKACPDRHMCQ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/JVM6ZPKQ4OZUBK2NKACPDRHMCQ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/JVM6ZPKQ4OZUBK2NKACPDRHMCQ/action/storage_attestation","attest_author":"https://pith.science/pith/JVM6ZPKQ4OZUBK2NKACPDRHMCQ/action/author_attestation","sign_citation":"https://pith.science/pith/JVM6ZPKQ4OZUBK2NKACPDRHMCQ/action/citation_signature","submit_replication":"https://pith.science/pith/JVM6ZPKQ4OZUBK2NKACPDRHMCQ/action/replication_record"}},"created_at":"2026-07-22T00:22:58.063386+00:00","updated_at":"2026-07-22T00:22:58.063386+00:00"}