{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:AX5VZLE7WIKRFR3BXXPZ7FPAZS","short_pith_number":"pith:AX5VZLE7","schema_version":"1.0","canonical_sha256":"05fb5cac9fb21512c761bddf9f95e0cc92492d3d377bed11d49018efb8c70d51","source":{"kind":"arxiv","id":"2608.10513","version":1},"attestation_state":"computed","paper":{"title":"SafeCap: Improving LVLM Safety with Image Captioning Reinforcement Learning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CV","authors_text":"Caoyuan Ma, Ji Zhao, Lingxi Min, Shilei Zhao, Shuai Dong, Taiqiang Wu, Tian Gu, Weichu Xie, Wenpu Liu, Wenqi Shao, Wenzheng Chang, Yinqiang Zheng, Yongfu Zhu, Yuqi Xu, Ziyue Wang","submitted_at":"2026-08-11T05:37:59Z","abstract_excerpt":"Large vision-language models (LVLMs) remain vulnerable to jailbreak attacks that exploit visual inputs to bypass safety alignment inherited from their language backbones. We propose SafeCap, a reinforcement-learning framework that aligns LVLMs through learned self-captioning. SafeCap trains a policy model to first generate a safety-relevant image caption and then produce a final answer; the caption is further optimized by whether it enables a frozen LLM to reach a safety-aligned decision. This caption-mediated objective encourages the policy to expose visual cues relevant to safe response gene"},"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":"2608.10513","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CV","submitted_at":"2026-08-11T05:37:59Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"cd193e7ab80dd938380e4f9d1b9df0e24e8df57dcdeeedf17108851a62db0620","abstract_canon_sha256":"ced1769daa9aa4318a0b1f90119b42eb06c683c4d774fde0a06383eafae039c3"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-08-12T01:22:38.783383Z","signature_b64":"sCD1G1MprMCGpimecTJ2DvBJax6KG+3t8X5KqG79A2ACRgkXqLEdlOQp+8bM4hP2MHAABfCgesgFlw9bBUrvDw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"05fb5cac9fb21512c761bddf9f95e0cc92492d3d377bed11d49018efb8c70d51","last_reissued_at":"2026-08-12T01:22:38.781813Z","signature_status":"signed_v1","first_computed_at":"2026-08-12T01:22:38.781813Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"SafeCap: Improving LVLM Safety with Image Captioning Reinforcement Learning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CV","authors_text":"Caoyuan Ma, Ji Zhao, Lingxi Min, Shilei Zhao, Shuai Dong, Taiqiang Wu, Tian Gu, Weichu Xie, Wenpu Liu, Wenqi Shao, Wenzheng Chang, Yinqiang Zheng, Yongfu Zhu, Yuqi Xu, Ziyue Wang","submitted_at":"2026-08-11T05:37:59Z","abstract_excerpt":"Large vision-language models (LVLMs) remain vulnerable to jailbreak attacks that exploit visual inputs to bypass safety alignment inherited from their language backbones. We propose SafeCap, a reinforcement-learning framework that aligns LVLMs through learned self-captioning. SafeCap trains a policy model to first generate a safety-relevant image caption and then produce a final answer; the caption is further optimized by whether it enables a frozen LLM to reach a safety-aligned decision. This caption-mediated objective encourages the policy to expose visual cues relevant to safe response gene"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2608.10513","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/2608.10513/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":"2608.10513","created_at":"2026-08-12T01:22:38.785637+00:00"},{"alias_kind":"arxiv_version","alias_value":"2608.10513v1","created_at":"2026-08-12T01:22:38.785637+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2608.10513","created_at":"2026-08-12T01:22:38.785637+00:00"},{"alias_kind":"pith_short_12","alias_value":"AX5VZLE7WIKR","created_at":"2026-08-12T01:22:38.785637+00:00"},{"alias_kind":"pith_short_16","alias_value":"AX5VZLE7WIKRFR3B","created_at":"2026-08-12T01:22:38.785637+00:00"},{"alias_kind":"pith_short_8","alias_value":"AX5VZLE7","created_at":"2026-08-12T01:22:38.785637+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/AX5VZLE7WIKRFR3BXXPZ7FPAZS","json":"https://pith.science/pith/AX5VZLE7WIKRFR3BXXPZ7FPAZS.json","graph_json":"https://pith.science/api/pith-number/AX5VZLE7WIKRFR3BXXPZ7FPAZS/graph.json","events_json":"https://pith.science/api/pith-number/AX5VZLE7WIKRFR3BXXPZ7FPAZS/events.json","paper":"https://pith.science/paper/AX5VZLE7"},"agent_actions":{"view_html":"https://pith.science/pith/AX5VZLE7WIKRFR3BXXPZ7FPAZS","download_json":"https://pith.science/pith/AX5VZLE7WIKRFR3BXXPZ7FPAZS.json","view_paper":"https://pith.science/paper/AX5VZLE7","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2608.10513&json=true","fetch_graph":"https://pith.science/api/pith-number/AX5VZLE7WIKRFR3BXXPZ7FPAZS/graph.json","fetch_events":"https://pith.science/api/pith-number/AX5VZLE7WIKRFR3BXXPZ7FPAZS/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/AX5VZLE7WIKRFR3BXXPZ7FPAZS/action/timestamp_anchor","attest_storage":"https://pith.science/pith/AX5VZLE7WIKRFR3BXXPZ7FPAZS/action/storage_attestation","attest_author":"https://pith.science/pith/AX5VZLE7WIKRFR3BXXPZ7FPAZS/action/author_attestation","sign_citation":"https://pith.science/pith/AX5VZLE7WIKRFR3BXXPZ7FPAZS/action/citation_signature","submit_replication":"https://pith.science/pith/AX5VZLE7WIKRFR3BXXPZ7FPAZS/action/replication_record"}},"created_at":"2026-08-12T01:22:38.785637+00:00","updated_at":"2026-08-12T01:22:38.785637+00:00"}