{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:LULA7QCB3UICUCDRVVQSQP2UVD","short_pith_number":"pith:LULA7QCB","schema_version":"1.0","canonical_sha256":"5d160fc041dd102a0871ad61283f54a8ff42da614d5632c518e09bcecc3688be","source":{"kind":"arxiv","id":"2509.00373","version":1},"attestation_state":"computed","paper":{"title":"Activation Steering Meets Preference Optimization: Defense Against Jailbreaks in Vision Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CV","authors_text":"Gaojie Jin, Jianhong Wang, Sihao Wu, Wei Huang, Xiaowei Huang","submitted_at":"2025-08-30T06:00:53Z","abstract_excerpt":"Vision Language Models (VLMs) have demonstrated impressive capabilities in integrating visual and textual information for understanding and reasoning, but remain highly vulnerable to adversarial attacks. While activation steering has emerged as a promising defence, existing approaches often rely on task-specific contrastive prompts to extract harmful directions, which exhibit suboptimal performance and can degrade visual grounding performance. To address these limitations, we propose \\textit{Sequence-Level Preference Optimization} for VLM (\\textit{SPO-VLM}), a novel two-stage defense framework"},"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":"2509.00373","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CV","submitted_at":"2025-08-30T06:00:53Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"a328aba3b234f21c8ace1941cf10950b685de1db5e35cb36e08afad68db88adf","abstract_canon_sha256":"1191cffe28603c4bab6b69e3b915f3e2211762ced7a9ec7e8dc1850cc0ea30c9"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T12:02:14.911303Z","signature_b64":"LyOIh1zJNJi5cXKXXk2pKG8zTKtgVy89O9CL3+K1D6tO99dLcEy7YR4rT+I1od9mpovPooKfio/ThQqTwQalCQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"5d160fc041dd102a0871ad61283f54a8ff42da614d5632c518e09bcecc3688be","last_reissued_at":"2026-07-05T12:02:14.910816Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T12:02:14.910816Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Activation Steering Meets Preference Optimization: Defense Against Jailbreaks in Vision Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CV","authors_text":"Gaojie Jin, Jianhong Wang, Sihao Wu, Wei Huang, Xiaowei Huang","submitted_at":"2025-08-30T06:00:53Z","abstract_excerpt":"Vision Language Models (VLMs) have demonstrated impressive capabilities in integrating visual and textual information for understanding and reasoning, but remain highly vulnerable to adversarial attacks. While activation steering has emerged as a promising defence, existing approaches often rely on task-specific contrastive prompts to extract harmful directions, which exhibit suboptimal performance and can degrade visual grounding performance. To address these limitations, we propose \\textit{Sequence-Level Preference Optimization} for VLM (\\textit{SPO-VLM}), a novel two-stage defense framework"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2509.00373","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/2509.00373/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":"2509.00373","created_at":"2026-07-05T12:02:14.910873+00:00"},{"alias_kind":"arxiv_version","alias_value":"2509.00373v1","created_at":"2026-07-05T12:02:14.910873+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2509.00373","created_at":"2026-07-05T12:02:14.910873+00:00"},{"alias_kind":"pith_short_12","alias_value":"LULA7QCB3UIC","created_at":"2026-07-05T12:02:14.910873+00:00"},{"alias_kind":"pith_short_16","alias_value":"LULA7QCB3UICUCDR","created_at":"2026-07-05T12:02:14.910873+00:00"},{"alias_kind":"pith_short_8","alias_value":"LULA7QCB","created_at":"2026-07-05T12:02:14.910873+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.15687","citing_title":"ASRU: Activation Steering Meets Reinforcement Unlearning for Multimodal Large Language Models","ref_index":24,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/LULA7QCB3UICUCDRVVQSQP2UVD","json":"https://pith.science/pith/LULA7QCB3UICUCDRVVQSQP2UVD.json","graph_json":"https://pith.science/api/pith-number/LULA7QCB3UICUCDRVVQSQP2UVD/graph.json","events_json":"https://pith.science/api/pith-number/LULA7QCB3UICUCDRVVQSQP2UVD/events.json","paper":"https://pith.science/paper/LULA7QCB"},"agent_actions":{"view_html":"https://pith.science/pith/LULA7QCB3UICUCDRVVQSQP2UVD","download_json":"https://pith.science/pith/LULA7QCB3UICUCDRVVQSQP2UVD.json","view_paper":"https://pith.science/paper/LULA7QCB","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2509.00373&json=true","fetch_graph":"https://pith.science/api/pith-number/LULA7QCB3UICUCDRVVQSQP2UVD/graph.json","fetch_events":"https://pith.science/api/pith-number/LULA7QCB3UICUCDRVVQSQP2UVD/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/LULA7QCB3UICUCDRVVQSQP2UVD/action/timestamp_anchor","attest_storage":"https://pith.science/pith/LULA7QCB3UICUCDRVVQSQP2UVD/action/storage_attestation","attest_author":"https://pith.science/pith/LULA7QCB3UICUCDRVVQSQP2UVD/action/author_attestation","sign_citation":"https://pith.science/pith/LULA7QCB3UICUCDRVVQSQP2UVD/action/citation_signature","submit_replication":"https://pith.science/pith/LULA7QCB3UICUCDRVVQSQP2UVD/action/replication_record"}},"created_at":"2026-07-05T12:02:14.910873+00:00","updated_at":"2026-07-05T12:02:14.910873+00:00"}