{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:VQMPFD7EHMUMO6FVYUPLBU6POS","short_pith_number":"pith:VQMPFD7E","schema_version":"1.0","canonical_sha256":"ac18f28fe43b28c778b5c51eb0d3cf74b15b8cf4701039e8417b656176d45bf0","source":{"kind":"arxiv","id":"2605.07919","version":2},"attestation_state":"computed","paper":{"title":"MedVIGIL: Evaluating Trustworthy Medical VLMs Under Broken Visual Evidence","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"Medical vision-language models fail to refuse answers when visual evidence is broken, trailing radiologists by 14 points on a new composite score.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Hanqi Jiang, Haozhen Gong, Hui Ren, Hyeokjae Kwon, Jinglei Lv, Junhao Chen, Lifeng Chen, Lin Zhao, Mingyu Kang, Quanzheng Li, Ruiyu Yan, Tianming Liu, Weihang You, Xiang Li, Yi Pan","submitted_at":"2026-05-08T15:55:30Z","abstract_excerpt":"Medical vision--language models (VLMs) are usually evaluated on intact image--question pairs, but trustworthy clinical use requires a stronger property: a model must recognise when the evidential basis for an answer has failed. We study this through silent failures under perturbed evidence, where a vision-required medical question is paired with a false premise, wording perturbation, knowledge-only rewrite, or ROI-corrupted image, yet the model returns a fluent non-refusal answer. We introduce medvigil, a 300-case evaluation suite drawn from four public medical VQA sources, supervised end to e"},"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":"2605.07919","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","primary_cat":"cs.CV","submitted_at":"2026-05-08T15:55:30Z","cross_cats_sorted":[],"title_canon_sha256":"53ac9af92f70e8a5751c4d6922585835fb90690dc00399e153d13ad726421550","abstract_canon_sha256":"133a9adea01e80b03ff81c4967815dba5865719a1437782d91a280f73ba236a3"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-25T02:02:16.509641Z","signature_b64":"vAlHsULuIJSPBomTOcKUcnXkUxCkUfgsQcWyKLCbEvTHKByTIlHvd0vHkhGqKT0K8tc9Usamc06CS1fzRGrVDA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"ac18f28fe43b28c778b5c51eb0d3cf74b15b8cf4701039e8417b656176d45bf0","last_reissued_at":"2026-05-25T02:02:16.508920Z","signature_status":"signed_v1","first_computed_at":"2026-05-25T02:02:16.508920Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"MedVIGIL: Evaluating Trustworthy Medical VLMs Under Broken Visual Evidence","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"Medical vision-language models fail to refuse answers when visual evidence is broken, trailing radiologists by 14 points on a new composite score.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Hanqi Jiang, Haozhen Gong, Hui Ren, Hyeokjae Kwon, Jinglei Lv, Junhao Chen, Lifeng Chen, Lin Zhao, Mingyu Kang, Quanzheng Li, Ruiyu Yan, Tianming Liu, Weihang You, Xiang Li, Yi Pan","submitted_at":"2026-05-08T15:55:30Z","abstract_excerpt":"Medical vision--language models (VLMs) are usually evaluated on intact image--question pairs, but trustworthy clinical use requires a stronger property: a model must recognise when the evidential basis for an answer has failed. We study this through silent failures under perturbed evidence, where a vision-required medical question is paired with a false premise, wording perturbation, knowledge-only rewrite, or ROI-corrupted image, yet the model returns a fluent non-refusal answer. We introduce medvigil, a 300-case evaluation suite drawn from four public medical VQA sources, supervised end to e"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"The independent radiologist scores MCS 83.3 at silent-failure rate 5.8%, leaving a 14.1-point composite headroom above the strongest audited model (Claude Opus 4.7 at 69.2).","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the four chosen perturbation types and the 300 clinician-authored cases sufficiently represent the range of broken visual evidence that occurs in real clinical practice.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"MedVIGIL introduces a clinician-supervised benchmark showing medical VLMs frequently give fluent answers on broken visual evidence, with top models 14 points below human radiologists on the composite score.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Medical vision-language models fail to refuse answers when visual evidence is broken, trailing radiologists by 14 points on a new composite score.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"5abf56945b417f28f53c287fd507a3ccfd3d0c308bedf4408748d780b5d4b2a2"},"source":{"id":"2605.07919","kind":"arxiv","version":2},"verdict":{"id":"40477877-05fe-4735-b79f-76bbcd099224","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-11T03:40:45.192215Z","strongest_claim":"The independent radiologist scores MCS 83.3 at silent-failure rate 5.8%, leaving a 14.1-point composite headroom above the strongest audited model (Claude Opus 4.7 at 69.2).","one_line_summary":"MedVIGIL introduces a clinician-supervised benchmark showing medical VLMs frequently give fluent answers on broken visual evidence, with top models 14 points below human radiologists on the composite score.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the four chosen perturbation types and the 300 clinician-authored cases sufficiently represent the range of broken visual evidence that occurs in real clinical practice.","pith_extraction_headline":"Medical vision-language models fail to refuse answers when visual evidence is broken, trailing radiologists by 14 points on a new composite score."},"integrity":{"clean":false,"summary":{"advisory":2,"critical":0,"by_detector":{"doi_compliance":{"total":2,"advisory":2,"critical":0,"informational":0}},"informational":0},"endpoint":"/pith/2605.07919/integrity.json","findings":[{"note":"DOI in the printed bibliography is fragmented by whitespace or line breaks. A longer candidate (10.1109/TBME.2021.3117407.URLhttps://pmc.ncbi.nlm.nih.gov/articles/PMC9011180/.Dan) was visible in the surrounding text but could not be confirmed against doi.org as printed.","detector":"doi_compliance","severity":"advisory","ref_index":10,"audited_at":"2026-05-19T11:24:20.741484Z","detected_doi":"10.1109/TBME.2021.3117407.URLhttps://pmc.ncbi.nlm.nih.gov/articles/PMC9011180/.Dan","finding_type":"recoverable_identifier","verdict_class":"incontrovertible","detected_arxiv_id":null},{"note":"DOI in the printed bibliography is fragmented by whitespace or line breaks. A longer candidate (10.18653/v1/2023.emnlp-main.20.URLhttps://aclanthology.org/2023.emnlp-main.20/.Bo) was visible in the surrounding text but could not be confirmed against doi.org as printed.","detector":"doi_compliance","severity":"advisory","ref_index":17,"audited_at":"2026-05-19T11:24:20.741484Z","detected_doi":"10.18653/v1/2023.emnlp-main.20.URLhttps://aclanthology.org/2023.emnlp-main.20/.Bo","finding_type":"recoverable_identifier","verdict_class":"incontrovertible","detected_arxiv_id":null}],"available":true,"detectors_run":[{"name":"claim_evidence","ran_at":"2026-05-20T10:02:13.555939Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-20T04:46:22.489233Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-19T15:31:18.310769Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T11:24:20.741484Z","status":"completed","version":"1.0.0","findings_count":2}],"snapshot_sha256":"4abdae5894b4b9e13c60c8601c856f602b2de3eb7fdd7894a72fce302b49f1cc"},"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":"2605.07919","created_at":"2026-05-25T02:02:16.509026+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.07919v2","created_at":"2026-05-25T02:02:16.509026+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.07919","created_at":"2026-05-25T02:02:16.509026+00:00"},{"alias_kind":"pith_short_12","alias_value":"VQMPFD7EHMUM","created_at":"2026-05-25T02:02:16.509026+00:00"},{"alias_kind":"pith_short_16","alias_value":"VQMPFD7EHMUMO6FV","created_at":"2026-05-25T02:02:16.509026+00:00"},{"alias_kind":"pith_short_8","alias_value":"VQMPFD7E","created_at":"2026-05-25T02:02:16.509026+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/VQMPFD7EHMUMO6FVYUPLBU6POS","json":"https://pith.science/pith/VQMPFD7EHMUMO6FVYUPLBU6POS.json","graph_json":"https://pith.science/api/pith-number/VQMPFD7EHMUMO6FVYUPLBU6POS/graph.json","events_json":"https://pith.science/api/pith-number/VQMPFD7EHMUMO6FVYUPLBU6POS/events.json","paper":"https://pith.science/paper/VQMPFD7E"},"agent_actions":{"view_html":"https://pith.science/pith/VQMPFD7EHMUMO6FVYUPLBU6POS","download_json":"https://pith.science/pith/VQMPFD7EHMUMO6FVYUPLBU6POS.json","view_paper":"https://pith.science/paper/VQMPFD7E","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.07919&json=true","fetch_graph":"https://pith.science/api/pith-number/VQMPFD7EHMUMO6FVYUPLBU6POS/graph.json","fetch_events":"https://pith.science/api/pith-number/VQMPFD7EHMUMO6FVYUPLBU6POS/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/VQMPFD7EHMUMO6FVYUPLBU6POS/action/timestamp_anchor","attest_storage":"https://pith.science/pith/VQMPFD7EHMUMO6FVYUPLBU6POS/action/storage_attestation","attest_author":"https://pith.science/pith/VQMPFD7EHMUMO6FVYUPLBU6POS/action/author_attestation","sign_citation":"https://pith.science/pith/VQMPFD7EHMUMO6FVYUPLBU6POS/action/citation_signature","submit_replication":"https://pith.science/pith/VQMPFD7EHMUMO6FVYUPLBU6POS/action/replication_record"}},"created_at":"2026-05-25T02:02:16.509026+00:00","updated_at":"2026-05-25T02:02:16.509026+00:00"}