Beyond the Evidence Boundary: Behavioral Mirage Under Progressive Degradation in Medical VLMs

Key points

• VLMs for medical image interpretation are typically evaluated on clean inputs, despite the progressive evidence loss that can occur in high stakes clinical settings.1
• MedGemma's post-boundary behavior is heterogeneous, including direct mirage, correct recovery, mirage dissolution, consistent abstention, and oscillation, with mirage rates of 16.0 to 100.0% across dataset-modality pairs.1
• Despite this behavioral heterogeneity, residual-stream probes decode commit-versus-abstain behavior at the boundary with AUROC of 0.895 to 0.994 under nested cross-validation.1
• These results show that progressive evidence loss can produce persistent commitment that output entropy and abstention behavior do not consistently reflect, motivating boundary-aware evaluation, option-order controls, and improved abstention mechanisms for reliable medical foundation models operating under partial observation.1

Sources (1)

• [1]Beyond the Evidence Boundary: Behavioral Mirage Under Progressive Degradation in Medical VLMs
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 20, 10:45 AM

  We introduce a baseline-anchored framework that traces baseline-correct cases through progressive image degradation to identify an empirical evidence boundary, and define mirage as continued commitment at this boundary, regardless of correctness.

  VLMs for medical image interpretation are typically evaluated on clean inputs, despite the progressive evidence loss that can occur in high stakes clinical settings.

Extractive summary: sentences quoted from the sources.