{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:F4CTTAVUN5UU4HXEQZAP7WRAYH","short_pith_number":"pith:F4CTTAVU","schema_version":"1.0","canonical_sha256":"2f053982b46f694e1ee48640ffda20c1eab39e804f2129ddb3ba0b481ae4f93e","source":{"kind":"arxiv","id":"2410.06625","version":2},"attestation_state":"computed","paper":{"title":"ETA: Evaluating Then Aligning Safety of Vision Language Models at Inference Time","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL","cs.LG"],"primary_cat":"cs.CV","authors_text":"Bolian Li, Ruqi Zhang, Yi Ding","submitted_at":"2024-10-09T07:21:43Z","abstract_excerpt":"Vision Language Models (VLMs) have become essential backbones for multimodal intelligence, yet significant safety challenges limit their real-world application. While textual inputs are often effectively safeguarded, adversarial visual inputs can easily bypass VLM defense mechanisms. Existing defense methods are either resource-intensive, requiring substantial data and compute, or fail to simultaneously ensure safety and usefulness in responses. To address these limitations, we propose a novel two-phase inference-time alignment framework, Evaluating Then Aligning (ETA): 1) Evaluating input vis"},"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":"2410.06625","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2024-10-09T07:21:43Z","cross_cats_sorted":["cs.CL","cs.LG"],"title_canon_sha256":"2a582818d691bc2099c005e7976b94712408678bb493ad0369e1e6577330e6c5","abstract_canon_sha256":"30045cca7967dc71187ed516370db9ef9855b6d80305827ffbd3996261f500a0"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:11:25.499091Z","signature_b64":"bn3biuwQwD66pgnJ4IG7ittdztCVZdHpuVDgd1hQSMTOYDBZjM8VeM/NRM3frQUadVqf3m72dA4XAgkZMXJ2Bw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"2f053982b46f694e1ee48640ffda20c1eab39e804f2129ddb3ba0b481ae4f93e","last_reissued_at":"2026-07-05T10:11:25.498196Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:11:25.498196Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"ETA: Evaluating Then Aligning Safety of Vision Language Models at Inference Time","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL","cs.LG"],"primary_cat":"cs.CV","authors_text":"Bolian Li, Ruqi Zhang, Yi Ding","submitted_at":"2024-10-09T07:21:43Z","abstract_excerpt":"Vision Language Models (VLMs) have become essential backbones for multimodal intelligence, yet significant safety challenges limit their real-world application. While textual inputs are often effectively safeguarded, adversarial visual inputs can easily bypass VLM defense mechanisms. Existing defense methods are either resource-intensive, requiring substantial data and compute, or fail to simultaneously ensure safety and usefulness in responses. To address these limitations, we propose a novel two-phase inference-time alignment framework, Evaluating Then Aligning (ETA): 1) Evaluating input vis"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2410.06625","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/2410.06625/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":"2410.06625","created_at":"2026-07-05T10:11:25.498250+00:00"},{"alias_kind":"arxiv_version","alias_value":"2410.06625v2","created_at":"2026-07-05T10:11:25.498250+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2410.06625","created_at":"2026-07-05T10:11:25.498250+00:00"},{"alias_kind":"pith_short_12","alias_value":"F4CTTAVUN5UU","created_at":"2026-07-05T10:11:25.498250+00:00"},{"alias_kind":"pith_short_16","alias_value":"F4CTTAVUN5UU4HXE","created_at":"2026-07-05T10:11:25.498250+00:00"},{"alias_kind":"pith_short_8","alias_value":"F4CTTAVU","created_at":"2026-07-05T10:11:25.498250+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.15030","citing_title":"WARD: Adversarially Robust Defense of Web Agents Against Prompt Injections","ref_index":18,"is_internal_anchor":false},{"citing_arxiv_id":"2604.08881","citing_title":"Targeted Interpretable Safety Neuron Enhancement for Multilingual Vision-Language Large Models","ref_index":5,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/F4CTTAVUN5UU4HXEQZAP7WRAYH","json":"https://pith.science/pith/F4CTTAVUN5UU4HXEQZAP7WRAYH.json","graph_json":"https://pith.science/api/pith-number/F4CTTAVUN5UU4HXEQZAP7WRAYH/graph.json","events_json":"https://pith.science/api/pith-number/F4CTTAVUN5UU4HXEQZAP7WRAYH/events.json","paper":"https://pith.science/paper/F4CTTAVU"},"agent_actions":{"view_html":"https://pith.science/pith/F4CTTAVUN5UU4HXEQZAP7WRAYH","download_json":"https://pith.science/pith/F4CTTAVUN5UU4HXEQZAP7WRAYH.json","view_paper":"https://pith.science/paper/F4CTTAVU","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2410.06625&json=true","fetch_graph":"https://pith.science/api/pith-number/F4CTTAVUN5UU4HXEQZAP7WRAYH/graph.json","fetch_events":"https://pith.science/api/pith-number/F4CTTAVUN5UU4HXEQZAP7WRAYH/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/F4CTTAVUN5UU4HXEQZAP7WRAYH/action/timestamp_anchor","attest_storage":"https://pith.science/pith/F4CTTAVUN5UU4HXEQZAP7WRAYH/action/storage_attestation","attest_author":"https://pith.science/pith/F4CTTAVUN5UU4HXEQZAP7WRAYH/action/author_attestation","sign_citation":"https://pith.science/pith/F4CTTAVUN5UU4HXEQZAP7WRAYH/action/citation_signature","submit_replication":"https://pith.science/pith/F4CTTAVUN5UU4HXEQZAP7WRAYH/action/replication_record"}},"created_at":"2026-07-05T10:11:25.498250+00:00","updated_at":"2026-07-05T10:11:25.498250+00:00"}