{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:XOKSKZUDK5E2WQUJLES6H5KQMD","short_pith_number":"pith:XOKSKZUD","schema_version":"1.0","canonical_sha256":"bb952566835749ab42895925e3f55060fc802615f06a191f30bd44b29c6c9fe2","source":{"kind":"arxiv","id":"2503.17682","version":2},"attestation_state":"computed","paper":{"title":"Safe RLHF-V: Safe Reinforcement Learning from Multi-modal Human Feedback","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Boyuan Chen, Chi-Min Chan, Conghui Zhang, Donghai Hong, Han Zhu, Jiahao Li, Jiaming Ji, Jiayi Zhou, Juntao Dai, Kaile Wang, Rui Pan, Sirui Han, Xinyu Chen, Yaodong Yang, Yida Tang, Yike Guo","submitted_at":"2025-03-22T07:40:20Z","abstract_excerpt":"Multimodal large language models (MLLMs) are essential for building general-purpose AI assistants; however, they pose increasing safety risks. How can we ensure safety alignment of MLLMs to prevent undesired behaviors? Going further, it is critical to explore how to fine-tune MLLMs to preserve capabilities while meeting safety constraints. Fundamentally, this challenge can be formulated as a min-max optimization problem. However, existing datasets have not yet disentangled single preference signals into explicit safety constraints, hindering systematic investigation in this direction. Moreover"},"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":"2503.17682","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2025-03-22T07:40:20Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"ad6d9aef8b7a7ffce83c8341dd48f091f5b8c1821ff705a1b18a4796345dae80","abstract_canon_sha256":"df59132b7dba76d37501a36a0a16895d54a15689052331b46d60544245a543f3"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:07:21.950938Z","signature_b64":"ZN+nockJIWHCJVG0o0s7TxBMn0+D3nXJ/hSqrBch6YZ8Cq6oL9KIDYhLmKUWS7uXg0xVz5j23pGM1/hho3/oBA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"bb952566835749ab42895925e3f55060fc802615f06a191f30bd44b29c6c9fe2","last_reissued_at":"2026-07-05T11:07:21.950383Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:07:21.950383Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Safe RLHF-V: Safe Reinforcement Learning from Multi-modal Human Feedback","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Boyuan Chen, Chi-Min Chan, Conghui Zhang, Donghai Hong, Han Zhu, Jiahao Li, Jiaming Ji, Jiayi Zhou, Juntao Dai, Kaile Wang, Rui Pan, Sirui Han, Xinyu Chen, Yaodong Yang, Yida Tang, Yike Guo","submitted_at":"2025-03-22T07:40:20Z","abstract_excerpt":"Multimodal large language models (MLLMs) are essential for building general-purpose AI assistants; however, they pose increasing safety risks. How can we ensure safety alignment of MLLMs to prevent undesired behaviors? Going further, it is critical to explore how to fine-tune MLLMs to preserve capabilities while meeting safety constraints. Fundamentally, this challenge can be formulated as a min-max optimization problem. However, existing datasets have not yet disentangled single preference signals into explicit safety constraints, hindering systematic investigation in this direction. Moreover"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2503.17682","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/2503.17682/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":"2503.17682","created_at":"2026-07-05T11:07:21.950461+00:00"},{"alias_kind":"arxiv_version","alias_value":"2503.17682v2","created_at":"2026-07-05T11:07:21.950461+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2503.17682","created_at":"2026-07-05T11:07:21.950461+00:00"},{"alias_kind":"pith_short_12","alias_value":"XOKSKZUDK5E2","created_at":"2026-07-05T11:07:21.950461+00:00"},{"alias_kind":"pith_short_16","alias_value":"XOKSKZUDK5E2WQUJ","created_at":"2026-07-05T11:07:21.950461+00:00"},{"alias_kind":"pith_short_8","alias_value":"XOKSKZUD","created_at":"2026-07-05T11:07:21.950461+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":6,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.15030","citing_title":"WARD: Adversarially Robust Defense of Web Agents Against Prompt Injections","ref_index":28,"is_internal_anchor":false},{"citing_arxiv_id":"2606.11953","citing_title":"Decoding Multimodal Cues: Unveiling the Implicit Meaning Behind Hateful Videos","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2503.03480","citing_title":"SafeVLA: Towards Safety Alignment of Vision-Language-Action Model via Constrained Learning","ref_index":53,"is_internal_anchor":false},{"citing_arxiv_id":"2605.18104","citing_title":"Safety Geometry Collapse in Multimodal LLMs and Adaptive Drift Correction","ref_index":55,"is_internal_anchor":false},{"citing_arxiv_id":"2605.04992","citing_title":"You Snooze, You Lose: Automatic Safety Alignment Restoration through Neural Weight Translation","ref_index":68,"is_internal_anchor":false},{"citing_arxiv_id":"2604.19024","citing_title":"Policy Gradient Primal-Dual Method for Safe Reinforcement Learning from Human Feedback","ref_index":50,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/XOKSKZUDK5E2WQUJLES6H5KQMD","json":"https://pith.science/pith/XOKSKZUDK5E2WQUJLES6H5KQMD.json","graph_json":"https://pith.science/api/pith-number/XOKSKZUDK5E2WQUJLES6H5KQMD/graph.json","events_json":"https://pith.science/api/pith-number/XOKSKZUDK5E2WQUJLES6H5KQMD/events.json","paper":"https://pith.science/paper/XOKSKZUD"},"agent_actions":{"view_html":"https://pith.science/pith/XOKSKZUDK5E2WQUJLES6H5KQMD","download_json":"https://pith.science/pith/XOKSKZUDK5E2WQUJLES6H5KQMD.json","view_paper":"https://pith.science/paper/XOKSKZUD","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2503.17682&json=true","fetch_graph":"https://pith.science/api/pith-number/XOKSKZUDK5E2WQUJLES6H5KQMD/graph.json","fetch_events":"https://pith.science/api/pith-number/XOKSKZUDK5E2WQUJLES6H5KQMD/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/XOKSKZUDK5E2WQUJLES6H5KQMD/action/timestamp_anchor","attest_storage":"https://pith.science/pith/XOKSKZUDK5E2WQUJLES6H5KQMD/action/storage_attestation","attest_author":"https://pith.science/pith/XOKSKZUDK5E2WQUJLES6H5KQMD/action/author_attestation","sign_citation":"https://pith.science/pith/XOKSKZUDK5E2WQUJLES6H5KQMD/action/citation_signature","submit_replication":"https://pith.science/pith/XOKSKZUDK5E2WQUJLES6H5KQMD/action/replication_record"}},"created_at":"2026-07-05T11:07:21.950461+00:00","updated_at":"2026-07-05T11:07:21.950461+00:00"}