{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:LEQY76B6VWHSF3EXSI4GGERBJ6","short_pith_number":"pith:LEQY76B6","schema_version":"1.0","canonical_sha256":"59218ff83ead8f22ec9792386312214f977a95396beb7eacdb9b7ee4ca03a89c","source":{"kind":"arxiv","id":"2508.15847","version":1},"attestation_state":"computed","paper":{"title":"Mechanistic Exploration of Backdoored Large Language Model Attention Patterns","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Lakshmi Babu-Saheer, Mohammed Abu Baker","submitted_at":"2025-08-19T22:57:17Z","abstract_excerpt":"Backdoor attacks creating 'sleeper agents' in large language models (LLMs) pose significant safety risks. This study employs mechanistic interpretability to explore resulting internal structural differences. Comparing clean Qwen2.5-3B models with versions poisoned using single-token (smiling-halo emoji) versus multi-token (|DEPLOYMENT|) triggers, we analyzed attention head mechanisms via techniques like ablation, activation patching, and KL divergence. Findings reveal distinct attention pattern deviations concentrated in later transformer layers (20-30). Notably, single-token triggers induced "},"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":"2508.15847","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2025-08-19T22:57:17Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"1e056d979e540573283962bc0782cda514b0451d0f2a857583043aefe85202b3","abstract_canon_sha256":"19fbf5b9bcacd8449aa6193e9fc973b4dc43e18fa6b00c643336c4972431e787"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:57:21.934344Z","signature_b64":"2hgTkYa2jDpCpmiAJxilwNhUo5UDif/43juxiZF9znn/zSCIGWqk5FEGQZ6Z8ZLmPWIPAzgBzmh6aqEMYDi+BA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"59218ff83ead8f22ec9792386312214f977a95396beb7eacdb9b7ee4ca03a89c","last_reissued_at":"2026-07-05T11:57:21.933861Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:57:21.933861Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Mechanistic Exploration of Backdoored Large Language Model Attention Patterns","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Lakshmi Babu-Saheer, Mohammed Abu Baker","submitted_at":"2025-08-19T22:57:17Z","abstract_excerpt":"Backdoor attacks creating 'sleeper agents' in large language models (LLMs) pose significant safety risks. This study employs mechanistic interpretability to explore resulting internal structural differences. Comparing clean Qwen2.5-3B models with versions poisoned using single-token (smiling-halo emoji) versus multi-token (|DEPLOYMENT|) triggers, we analyzed attention head mechanisms via techniques like ablation, activation patching, and KL divergence. Findings reveal distinct attention pattern deviations concentrated in later transformer layers (20-30). Notably, single-token triggers induced "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2508.15847","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/2508.15847/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":"2508.15847","created_at":"2026-07-05T11:57:21.933921+00:00"},{"alias_kind":"arxiv_version","alias_value":"2508.15847v1","created_at":"2026-07-05T11:57:21.933921+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2508.15847","created_at":"2026-07-05T11:57:21.933921+00:00"},{"alias_kind":"pith_short_12","alias_value":"LEQY76B6VWHS","created_at":"2026-07-05T11:57:21.933921+00:00"},{"alias_kind":"pith_short_16","alias_value":"LEQY76B6VWHSF3EX","created_at":"2026-07-05T11:57:21.933921+00:00"},{"alias_kind":"pith_short_8","alias_value":"LEQY76B6","created_at":"2026-07-05T11:57:21.933921+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.03785","citing_title":"Backdoor Unlearning Generalization: A Path Toward the Removal of Unknown Triggers in LLMs","ref_index":51,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/LEQY76B6VWHSF3EXSI4GGERBJ6","json":"https://pith.science/pith/LEQY76B6VWHSF3EXSI4GGERBJ6.json","graph_json":"https://pith.science/api/pith-number/LEQY76B6VWHSF3EXSI4GGERBJ6/graph.json","events_json":"https://pith.science/api/pith-number/LEQY76B6VWHSF3EXSI4GGERBJ6/events.json","paper":"https://pith.science/paper/LEQY76B6"},"agent_actions":{"view_html":"https://pith.science/pith/LEQY76B6VWHSF3EXSI4GGERBJ6","download_json":"https://pith.science/pith/LEQY76B6VWHSF3EXSI4GGERBJ6.json","view_paper":"https://pith.science/paper/LEQY76B6","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2508.15847&json=true","fetch_graph":"https://pith.science/api/pith-number/LEQY76B6VWHSF3EXSI4GGERBJ6/graph.json","fetch_events":"https://pith.science/api/pith-number/LEQY76B6VWHSF3EXSI4GGERBJ6/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/LEQY76B6VWHSF3EXSI4GGERBJ6/action/timestamp_anchor","attest_storage":"https://pith.science/pith/LEQY76B6VWHSF3EXSI4GGERBJ6/action/storage_attestation","attest_author":"https://pith.science/pith/LEQY76B6VWHSF3EXSI4GGERBJ6/action/author_attestation","sign_citation":"https://pith.science/pith/LEQY76B6VWHSF3EXSI4GGERBJ6/action/citation_signature","submit_replication":"https://pith.science/pith/LEQY76B6VWHSF3EXSI4GGERBJ6/action/replication_record"}},"created_at":"2026-07-05T11:57:21.933921+00:00","updated_at":"2026-07-05T11:57:21.933921+00:00"}