{
  "id": 7040,
  "url": "https://arxiv.org/abs/2603.18373v4",
  "title": "To See or To Please: Uncovering Visual Sycophancy and Split Beliefs in VLMs",
  "summary": "When VLMs answer correctly, do they genuinely rely on visual information? We introduce a Tri-Layer Diagnostic Framework with three per-sample metrics: Latent Anomaly Detection, Visual Necessity Score, and Competition Score, which disentangle perception, dependency, and alignment failures. Across 9 VLMs and 9,000 model-sample pairs under counterfactual blind, noise, and conflict interventions, 72.9% of samples exhibit Visual Sycophancy, a Split Beliefs pattern in which internal evidence is preser",
  "authors": "Rui Hong, Shuxue Quan",
  "category": "research",
  "topics": "safety-alignment,healthcare",
  "orgs": null,
  "regions": null,
  "published_at": "2026-03-19T00:15:05.000Z",
  "fetched_at": "2026-07-14T16:32:54.536Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/7040",
  "original_url": "https://arxiv.org/abs/2603.18373v4",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}