{"paper":{"title":"Benchmarking and Mitigating Sycophancy in Medical Vision Language Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Medical vision language models exhibit sycophancy driven by visual cues and authority signals, which a filtering strategy called VIPER can reduce.","cross_cats":["cs.AI"],"primary_cat":"cs.CV","authors_text":"Di Wang, Hongbin Lin, Jingwei Lv, Juangui Xu, Jun Wen, Lijie Hu, Shu Yang, Xinyue Xu, Zikun Guo","submitted_at":"2025-09-26T07:02:22Z","abstract_excerpt":"Visual language models (VLMs) have the potential to transform medical workflows. However, the deployment is limited by sycophancy. Despite this serious threat to patient safety, a systematic benchmark remains lacking. This paper addresses this gap by introducing a Medical benchmark that applies multiple templates to VLMs in a hierarchical medical visual question answering task. We find that current VLMs are highly susceptible to visual cues, with failure rates showing a correlation to model size or overall accuracy. we discover that perceived authority and user mimicry are powerful triggers, s"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Current VLMs are highly susceptible to visual cues, with failure rates showing a correlation to model size or overall accuracy; perceived authority and user mimicry are powerful triggers suggesting a bias mechanism independent of visual data; VIPER reduces sycophancy while maintaining interpretability and consistently outperforms baseline methods.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The hierarchical medical visual question answering templates and authority/mimicry triggers accurately capture real-world sycophancy without introducing artificial biases that would not appear in actual clinical interactions (stated in the abstract description of the benchmark construction).","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Introduces a medical sycophancy benchmark for VLMs and the VIPER strategy to reduce agreement with non-evidence cues while preserving interpretability.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Medical vision language models exhibit sycophancy driven by visual cues and authority signals, which a filtering strategy called VIPER can reduce.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"a5caaf94286556142a2ac39956165aef603d47cb41dd5eb1ce9a7cb300fde720"},"source":{"id":"2509.21979","kind":"arxiv","version":6},"verdict":{"id":"e82a390c-8ed6-4f54-9a93-c240b871bc79","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-18T13:58:53.537348Z","strongest_claim":"Current VLMs are highly susceptible to visual cues, with failure rates showing a correlation to model size or overall accuracy; perceived authority and user mimicry are powerful triggers suggesting a bias mechanism independent of visual data; VIPER reduces sycophancy while maintaining interpretability and consistently outperforms baseline methods.","one_line_summary":"Introduces a medical sycophancy benchmark for VLMs and the VIPER strategy to reduce agreement with non-evidence cues while preserving interpretability.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The hierarchical medical visual question answering templates and authority/mimicry triggers accurately capture real-world sycophancy without introducing artificial biases that would not appear in actual clinical interactions (stated in the abstract description of the benchmark construction).","pith_extraction_headline":"Medical vision language models exhibit sycophancy driven by visual cues and authority signals, which a filtering strategy called VIPER can reduce."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2509.21979/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":2,"snapshot_sha256":"e67f19ec4f8d73d626a689b4f1ed1f43ccb77231bd1b3d5840f8aa1ddabb51f1"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}