{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:UU5P7VXZW3ELMLIJKZVIWGW5OP","short_pith_number":"pith:UU5P7VXZ","schema_version":"1.0","canonical_sha256":"a53affd6f9b6c8b62d09566a8b1add73d493f44773463fe6aa81fda944622853","source":{"kind":"arxiv","id":"2502.13946","version":2},"attestation_state":"computed","paper":{"title":"Why Safeguarded Ships Run Aground? Aligned Large Language Models' Safety Mechanisms Tend to Be Anchored in The Template Region","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"","cross_cats":["cs.AI","cs.CR"],"primary_cat":"cs.CL","authors_text":"Chak Tou Leong, Jian Wang, Qingyu Yin, Wenjie Li","submitted_at":"2025-02-19T18:42:45Z","abstract_excerpt":"The safety alignment of large language models (LLMs) remains vulnerable, as their initial behavior can be easily jailbroken by even relatively simple attacks. Since infilling a fixed template between the input instruction and initial model output is a common practice for existing LLMs, we hypothesize that this template is a key factor behind their vulnerabilities: LLMs' safety-related decision-making overly relies on the aggregated information from the template region, which largely influences these models' safety behavior. We refer to this issue as template-anchored safety alignment. In this "},"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":"2502.13946","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","primary_cat":"cs.CL","submitted_at":"2025-02-19T18:42:45Z","cross_cats_sorted":["cs.AI","cs.CR"],"title_canon_sha256":"ace8f75482169593f58a6173ec427bde2e603ccda0c943457518fbacbc39e3cd","abstract_canon_sha256":"a4537ec332ad83dca51aedfc55e31d0db8109768884c9af6bd86c1a306ff3be5"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:15:19.491743Z","signature_b64":"RLKG7a4KC3TF+w3aBbpIK56FLkbnsp5UlSvRDKgTeTsuNLJcqE6Axu1iVUyQpqztgL+wFIeEp1wGUo0bB4fNDw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"a53affd6f9b6c8b62d09566a8b1add73d493f44773463fe6aa81fda944622853","last_reissued_at":"2026-07-05T11:15:19.491303Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:15:19.491303Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Why Safeguarded Ships Run Aground? Aligned Large Language Models' Safety Mechanisms Tend to Be Anchored in The Template Region","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"","cross_cats":["cs.AI","cs.CR"],"primary_cat":"cs.CL","authors_text":"Chak Tou Leong, Jian Wang, Qingyu Yin, Wenjie Li","submitted_at":"2025-02-19T18:42:45Z","abstract_excerpt":"The safety alignment of large language models (LLMs) remains vulnerable, as their initial behavior can be easily jailbroken by even relatively simple attacks. Since infilling a fixed template between the input instruction and initial model output is a common practice for existing LLMs, we hypothesize that this template is a key factor behind their vulnerabilities: LLMs' safety-related decision-making overly relies on the aggregated information from the template region, which largely influences these models' safety behavior. We refer to this issue as template-anchored safety alignment. In this "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2502.13946","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/2502.13946/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":"2502.13946","created_at":"2026-07-05T11:15:19.491374+00:00"},{"alias_kind":"arxiv_version","alias_value":"2502.13946v2","created_at":"2026-07-05T11:15:19.491374+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2502.13946","created_at":"2026-07-05T11:15:19.491374+00:00"},{"alias_kind":"pith_short_12","alias_value":"UU5P7VXZW3EL","created_at":"2026-07-05T11:15:19.491374+00:00"},{"alias_kind":"pith_short_16","alias_value":"UU5P7VXZW3ELMLIJ","created_at":"2026-07-05T11:15:19.491374+00:00"},{"alias_kind":"pith_short_8","alias_value":"UU5P7VXZ","created_at":"2026-07-05T11:15:19.491374+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2507.11878","citing_title":"LLMs Encode Harmfulness and Refusal Separately","ref_index":11,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/UU5P7VXZW3ELMLIJKZVIWGW5OP","json":"https://pith.science/pith/UU5P7VXZW3ELMLIJKZVIWGW5OP.json","graph_json":"https://pith.science/api/pith-number/UU5P7VXZW3ELMLIJKZVIWGW5OP/graph.json","events_json":"https://pith.science/api/pith-number/UU5P7VXZW3ELMLIJKZVIWGW5OP/events.json","paper":"https://pith.science/paper/UU5P7VXZ"},"agent_actions":{"view_html":"https://pith.science/pith/UU5P7VXZW3ELMLIJKZVIWGW5OP","download_json":"https://pith.science/pith/UU5P7VXZW3ELMLIJKZVIWGW5OP.json","view_paper":"https://pith.science/paper/UU5P7VXZ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2502.13946&json=true","fetch_graph":"https://pith.science/api/pith-number/UU5P7VXZW3ELMLIJKZVIWGW5OP/graph.json","fetch_events":"https://pith.science/api/pith-number/UU5P7VXZW3ELMLIJKZVIWGW5OP/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/UU5P7VXZW3ELMLIJKZVIWGW5OP/action/timestamp_anchor","attest_storage":"https://pith.science/pith/UU5P7VXZW3ELMLIJKZVIWGW5OP/action/storage_attestation","attest_author":"https://pith.science/pith/UU5P7VXZW3ELMLIJKZVIWGW5OP/action/author_attestation","sign_citation":"https://pith.science/pith/UU5P7VXZW3ELMLIJKZVIWGW5OP/action/citation_signature","submit_replication":"https://pith.science/pith/UU5P7VXZW3ELMLIJKZVIWGW5OP/action/replication_record"}},"created_at":"2026-07-05T11:15:19.491374+00:00","updated_at":"2026-07-05T11:15:19.491374+00:00"}