{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:3LETKX3DNAUKQ7ARXJYI56JAAQ","short_pith_number":"pith:3LETKX3D","schema_version":"1.0","canonical_sha256":"dac9355f636828a87c11ba708ef920042c6674e3d8eb0551a281e4439c5221c2","source":{"kind":"arxiv","id":"2412.18826","version":1},"attestation_state":"computed","paper":{"title":"RapGuard: Safeguarding Multimodal Large Language Models via Rationale-aware Defensive Prompting","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Xiangyu Yue, Yilei Jiang, Yingshui Tan","submitted_at":"2024-12-25T08:31:53Z","abstract_excerpt":"While Multimodal Large Language Models (MLLMs) have made remarkable progress in vision-language reasoning, they are also more susceptible to producing harmful content compared to models that focus solely on text. Existing defensive prompting techniques rely on a static, unified safety guideline that fails to account for the specific risks inherent in different multimodal contexts. To address these limitations, we propose RapGuard, a novel framework that uses multimodal chain-of-thought reasoning to dynamically generate scenario-specific safety prompts. RapGuard enhances safety by adapting its "},"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":"2412.18826","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2024-12-25T08:31:53Z","cross_cats_sorted":[],"title_canon_sha256":"1c475d327b5e4ef37b929b302e6de98999a82a09dc08e94ce649d7c87228ddf6","abstract_canon_sha256":"b89a8c247d8947dd61af2466a3357e2f6c9bebbd2ce8265fe7791e4da723b23e"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:54:11.414129Z","signature_b64":"m9WubmL7hN62uV28r2iCJlgluqh6f8dKEyAD4ow6gBY1mxDurFs9tyDHFcLVtyYZhv57TP4/DMXFFzp+pjjsCQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"dac9355f636828a87c11ba708ef920042c6674e3d8eb0551a281e4439c5221c2","last_reissued_at":"2026-07-05T09:54:11.413695Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:54:11.413695Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"RapGuard: Safeguarding Multimodal Large Language Models via Rationale-aware Defensive Prompting","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Xiangyu Yue, Yilei Jiang, Yingshui Tan","submitted_at":"2024-12-25T08:31:53Z","abstract_excerpt":"While Multimodal Large Language Models (MLLMs) have made remarkable progress in vision-language reasoning, they are also more susceptible to producing harmful content compared to models that focus solely on text. Existing defensive prompting techniques rely on a static, unified safety guideline that fails to account for the specific risks inherent in different multimodal contexts. To address these limitations, we propose RapGuard, a novel framework that uses multimodal chain-of-thought reasoning to dynamically generate scenario-specific safety prompts. RapGuard enhances safety by adapting its "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2412.18826","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/2412.18826/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":"2412.18826","created_at":"2026-07-05T09:54:11.413752+00:00"},{"alias_kind":"arxiv_version","alias_value":"2412.18826v1","created_at":"2026-07-05T09:54:11.413752+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2412.18826","created_at":"2026-07-05T09:54:11.413752+00:00"},{"alias_kind":"pith_short_12","alias_value":"3LETKX3DNAUK","created_at":"2026-07-05T09:54:11.413752+00:00"},{"alias_kind":"pith_short_16","alias_value":"3LETKX3DNAUKQ7AR","created_at":"2026-07-05T09:54:11.413752+00:00"},{"alias_kind":"pith_short_8","alias_value":"3LETKX3D","created_at":"2026-07-05T09:54:11.413752+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.10904","citing_title":"Auditing Inference-Time Defense Evaluation for Multimodal Large Language Models","ref_index":10,"is_internal_anchor":false},{"citing_arxiv_id":"2604.08846","citing_title":"Dictionary-Aligned Concept Control for Safeguarding Multimodal LLMs","ref_index":31,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/3LETKX3DNAUKQ7ARXJYI56JAAQ","json":"https://pith.science/pith/3LETKX3DNAUKQ7ARXJYI56JAAQ.json","graph_json":"https://pith.science/api/pith-number/3LETKX3DNAUKQ7ARXJYI56JAAQ/graph.json","events_json":"https://pith.science/api/pith-number/3LETKX3DNAUKQ7ARXJYI56JAAQ/events.json","paper":"https://pith.science/paper/3LETKX3D"},"agent_actions":{"view_html":"https://pith.science/pith/3LETKX3DNAUKQ7ARXJYI56JAAQ","download_json":"https://pith.science/pith/3LETKX3DNAUKQ7ARXJYI56JAAQ.json","view_paper":"https://pith.science/paper/3LETKX3D","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2412.18826&json=true","fetch_graph":"https://pith.science/api/pith-number/3LETKX3DNAUKQ7ARXJYI56JAAQ/graph.json","fetch_events":"https://pith.science/api/pith-number/3LETKX3DNAUKQ7ARXJYI56JAAQ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/3LETKX3DNAUKQ7ARXJYI56JAAQ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/3LETKX3DNAUKQ7ARXJYI56JAAQ/action/storage_attestation","attest_author":"https://pith.science/pith/3LETKX3DNAUKQ7ARXJYI56JAAQ/action/author_attestation","sign_citation":"https://pith.science/pith/3LETKX3DNAUKQ7ARXJYI56JAAQ/action/citation_signature","submit_replication":"https://pith.science/pith/3LETKX3DNAUKQ7ARXJYI56JAAQ/action/replication_record"}},"created_at":"2026-07-05T09:54:11.413752+00:00","updated_at":"2026-07-05T09:54:11.413752+00:00"}