{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:6HZGICBEC4BDLRXZ53YLCRDFYU","short_pith_number":"pith:6HZGICBE","schema_version":"1.0","canonical_sha256":"f1f2640824170235c6f9eef0b14465c535e3d0063c2291ded20deaffdf51854c","source":{"kind":"arxiv","id":"2411.12768","version":2},"attestation_state":"computed","paper":{"title":"CROW: Eliminating Backdoors from Large Language Models via Internal Consistency Regularization","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Jun Sun, Long H. Pham, Nay Myat Min, Yige Li","submitted_at":"2024-11-18T07:52:12Z","abstract_excerpt":"Large Language Models (LLMs) are vulnerable to backdoor attacks that manipulate outputs via hidden triggers. Existing defense methods--designed for vision/text classification tasks--fail for text generation. We propose Internal Consistency Regularization (CROW), a defense leveraging the observation that backdoored models exhibit unstable layer-wise hidden representations when triggered, while clean models show smooth transitions. CROW enforces consistency across layers via adversarial perturbations and regularization during finetuning, neutralizing backdoors without requiring clean reference m"},"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":"2411.12768","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2024-11-18T07:52:12Z","cross_cats_sorted":["cs.AI","cs.LG"],"title_canon_sha256":"40688d25904492a540bbc3396519f1f3a477b406fb512071c5cd7eb27fd4f626","abstract_canon_sha256":"47569937c4096e6e98f0063010dc1002ca8e8d4f6bd7458f3f461e1ad72e185f"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:19:50.084226Z","signature_b64":"YmDY2TyVad7CzVUtanMai3rAJgrNkbRsaGbdkddt/66NHGsPWzXqBPQkak5EmcPu0UqDbO9EVP1kmVacJpz6BA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"f1f2640824170235c6f9eef0b14465c535e3d0063c2291ded20deaffdf51854c","last_reissued_at":"2026-07-05T11:19:50.083812Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:19:50.083812Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"CROW: Eliminating Backdoors from Large Language Models via Internal Consistency Regularization","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Jun Sun, Long H. Pham, Nay Myat Min, Yige Li","submitted_at":"2024-11-18T07:52:12Z","abstract_excerpt":"Large Language Models (LLMs) are vulnerable to backdoor attacks that manipulate outputs via hidden triggers. Existing defense methods--designed for vision/text classification tasks--fail for text generation. We propose Internal Consistency Regularization (CROW), a defense leveraging the observation that backdoored models exhibit unstable layer-wise hidden representations when triggered, while clean models show smooth transitions. CROW enforces consistency across layers via adversarial perturbations and regularization during finetuning, neutralizing backdoors without requiring clean reference m"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2411.12768","kind":"arxiv","version":2},"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/2411.12768/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":"2411.12768","created_at":"2026-07-05T11:19:50.083868+00:00"},{"alias_kind":"arxiv_version","alias_value":"2411.12768v2","created_at":"2026-07-05T11:19:50.083868+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2411.12768","created_at":"2026-07-05T11:19:50.083868+00:00"},{"alias_kind":"pith_short_12","alias_value":"6HZGICBEC4BD","created_at":"2026-07-05T11:19:50.083868+00:00"},{"alias_kind":"pith_short_16","alias_value":"6HZGICBEC4BDLRXZ","created_at":"2026-07-05T11:19:50.083868+00:00"},{"alias_kind":"pith_short_8","alias_value":"6HZGICBE","created_at":"2026-07-05T11:19:50.083868+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":3,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2502.05206","citing_title":"Safety at Scale: A Comprehensive Survey of Large Model and Agent Safety","ref_index":172,"is_internal_anchor":false},{"citing_arxiv_id":"2604.24162","citing_title":"Defusing the Trigger: Plug-and-Play Defense for Backdoored LLMs via Tail-Risk Intrinsic Geometric Smoothing","ref_index":24,"is_internal_anchor":false},{"citing_arxiv_id":"2604.09748","citing_title":"Backdoors in RLVR: Jailbreak Backdoors in LLMs From Verifiable Reward","ref_index":6,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/6HZGICBEC4BDLRXZ53YLCRDFYU","json":"https://pith.science/pith/6HZGICBEC4BDLRXZ53YLCRDFYU.json","graph_json":"https://pith.science/api/pith-number/6HZGICBEC4BDLRXZ53YLCRDFYU/graph.json","events_json":"https://pith.science/api/pith-number/6HZGICBEC4BDLRXZ53YLCRDFYU/events.json","paper":"https://pith.science/paper/6HZGICBE"},"agent_actions":{"view_html":"https://pith.science/pith/6HZGICBEC4BDLRXZ53YLCRDFYU","download_json":"https://pith.science/pith/6HZGICBEC4BDLRXZ53YLCRDFYU.json","view_paper":"https://pith.science/paper/6HZGICBE","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2411.12768&json=true","fetch_graph":"https://pith.science/api/pith-number/6HZGICBEC4BDLRXZ53YLCRDFYU/graph.json","fetch_events":"https://pith.science/api/pith-number/6HZGICBEC4BDLRXZ53YLCRDFYU/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/6HZGICBEC4BDLRXZ53YLCRDFYU/action/timestamp_anchor","attest_storage":"https://pith.science/pith/6HZGICBEC4BDLRXZ53YLCRDFYU/action/storage_attestation","attest_author":"https://pith.science/pith/6HZGICBEC4BDLRXZ53YLCRDFYU/action/author_attestation","sign_citation":"https://pith.science/pith/6HZGICBEC4BDLRXZ53YLCRDFYU/action/citation_signature","submit_replication":"https://pith.science/pith/6HZGICBEC4BDLRXZ53YLCRDFYU/action/replication_record"}},"created_at":"2026-07-05T11:19:50.083868+00:00","updated_at":"2026-07-05T11:19:50.083868+00:00"}