A Benchmark for Hallucination Detection in VLMs for Gastrointestinal Endoscopy

2026-06-23 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Health & Medical Research, Medical Devices & Health Technology · Depth: Expert, quick

Summary

A new benchmark study addresses the critical issue of hallucination detection in Vision-Language Models (VLMs) for gastrointestinal (GI) endoscopy, an area largely underexplored compared to radiology. Researchers evaluated nine hallucination detection methods across three categories (black-box, gray-box, and white-box) on the Gut-VLM dataset, which comprises 4,392 test VQA pairs. Five VLMs were tested: MedGemma-4B, MedGemma-27B, LLaVA-Med-7B, LLaVA-v1.6-7B, and Lingshu-32B. The white-box method, ReXTrust, consistently achieved the highest AUC, peaking at 93.0 on MedGemma-4B, and demonstrated a statistically significant advantage of 19.5 AUC points on average over other methods. Token-level gray-box statistics (MaxEnt, MaxProb) were identified as the strongest non-white-box alternatives. The study also highlighted "confident confabulation" as a systemic failure mode for existing detection approaches.

Key takeaway

For AI scientists deploying VLMs in clinical settings, particularly for GI endoscopy, you should prioritize models that allow white-box access for hallucination detection. Methods like ReXTrust offer significantly higher accuracy, with an average 19.5 AUC point advantage over alternatives. If white-box access is not feasible, explore token-level gray-box statistics such as MaxEnt or MaxProb as the next best option, but remain vigilant for "confident confabulation" as a persistent failure mode.

Key insights

White-box methods offer a significant advantage for detecting hallucinations in VLMs used for gastrointestinal endoscopy.

Principles

• White-box access consistently improves VLM hallucination detection.
• Token-level gray-box statistics are strong non-white-box alternatives.
• Confident confabulation is a systemic VLM failure mode.

Method

Benchmarking nine hallucination detection methods (black-box, gray-box, white-box) on the Gut-VLM dataset (4,392 VQA pairs) across five diverse VLMs to evaluate performance and identify failure modes.

In practice

• Prioritize VLMs allowing white-box access for clinical deployment.
• Consider token-level gray-box methods if white-box access is unavailable.
• Be aware of "confident confabulation" in VLM outputs.

Topics

• VLM Hallucination Detection
• Gastrointestinal Endoscopy
• Clinical AI
• ReXTrust
• Model Benchmarking
• Vision-Language Models

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• CuraView: A Multi-Agent Framework for Medical Hallucination Detection with GraphRAG-Enhanced Knowledge Verification — CuraView uses GraphRAG and multi-agent verification to detect and explain medical LLM hallucinations, improving factual reliability.
• Vision-Language Models as Zero-Annotation Oracles in Histopathology — General-purpose Vision-Language Models (VLMs) offer robust, zero-annotation foreground segmentation in histopathology, generalizing across diverse stains.
• MedBench v5: A Dynamic, Process-Oriented, and Hallucination-Aware Benchmark for Clinical Multimodal Models — MedBench v5 evaluates clinical AI models dynamically, focusing on process stability and hallucination propagation under various information stressors.
• Steering the Verifiability of Multimodal AI Hallucinations — Multimodal AI hallucinations vary in human verifiability, requiring distinct intervention strategies for obvious versus elusive errors.
• Hallucination Detection and Correction in Medical VLMs via Counter-Evidence Verification — CoEV verifies VLM outputs against visual evidence to detect and correct medical hallucinations.
• TriLens: Per-Layer Logit-Lens Entropy for White-Box Hallucination Detection — Internal model uncertainty, specifically per-layer logit-lens entropy, can effectively signal hallucination before final output.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.