HEDGE: Hallucination Estimation via Dense Geometric Entropy for VQA with Vision-Language Models

TL;DR

The paper addresses hallucinations in vision-language models for medical VQA by introducing HEDGE, a modular framework that quantifies hallucination risk through dense geometric entropy under controlled visual perturbations. It combines sampling, two semantic clustering strategies (NLI-based and embedding-based), and geometry-based metrics (SE, RadFlag, VASE) in a domain-agnostic pipeline, evaluated on VQA-RAD and KvasirVQA-x1 across three architectures. Key findings show that model architecture, prompt design, and sampling budget jointly govern detectability, with VASE providing the most robust signal in most regimes and clustering choices offering complementary strengths. The work provides a reproducible benchmarking toolkit (hedge-bench) and pragmatic guidance for deploying hallucination detection in multimodal systems and for shaping future benchmarks.

Abstract

Vision-language models (VLMs) enable open-ended visual question answering but remain prone to hallucinations. We present HEDGE, a unified framework for hallucination detection that combines controlled visual perturbations, semantic clustering, and robust uncertainty metrics. HEDGE integrates sampling, distortion synthesis, clustering (entailment- and embedding-based), and metric computation into a reproducible pipeline applicable across multimodal architectures. Evaluations on VQA-RAD and KvasirVQA-x1 with three representative VLMs (LLaVA-Med, Med-Gemma, Qwen2.5-VL) reveal clear architecture- and prompt-dependent trends. Hallucination detectability is highest for unified-fusion models with dense visual tokenization (Qwen2.5-VL) and lowest for architectures with restricted tokenization (Med-Gemma). Embedding-based clustering often yields stronger separation when applied directly to the generated answers, whereas NLI-based clustering remains advantageous for LLaVA-Med and for longer, sentence-level responses. Across configurations, the VASE metric consistently provides the most robust hallucination signal, especially when paired with embedding clustering and a moderate sampling budget (n ~ 10-15). Prompt design also matters: concise, label-style outputs offer clearer semantic structure than syntactically constrained one-sentence responses. By framing hallucination detection as a geometric robustness problem shaped jointly by sampling scale, prompt structure, model architecture, and clustering strategy, HEDGE provides a principled, compute-aware foundation for evaluating multimodal reliability. The hedge-bench PyPI library enables reproducible and extensible benchmarking, with full code and experimental resources available at https://github.com/Simula/HEDGE .

Figures (3)

• Figure 1: Overview of the proposed framework for hallucination detection in visual question answering (VQA). A vision--language model generates multiple answers per image--question pair, which are grouped via two strategies: natural language inference (NLI)-based logical clustering and embedding-aligned semantic clustering. Entropy within these groups quantifies uncertainty, enabling hallucination detection through metrics such as RadFlag, Semantic Entropy, and VASE.
• Figure 2: Effect of sampling scale and the associated visual perturbations on hallucination-detection performance for Qwen2.5-VL-7B-Instruct evaluated on the VQA-RAD dataset. ROC AUC scores are reported for three methods—Semantic Entropy (SE), RadFlag, and Vision-Amplified Semantic Entropy (VASE)—evaluated across four answer-length prompt configurations, including minimal-label, clinical-phrase, default, and one-sentence. Solid lines denote embedding-based clustering, whereas dashed lines denote NLI-based clustering.
• Figure 3: Scalability analysis of computational resources for NLI-based and embedding-based hallucination detection, using Qwen2.5-VL-7B-Instruct on the VQA-RAD dataset. (a) Processing time (logarithmic scale) and (b) NLI-related GPU memory usage, as a function of the maximum number of visual distortions (sampling scale). While panel (a) shows the NLI-based approach scales heavily in time, panel (b) focuses only on its GPU cost. The embedding-based clustering is ultra-light in comparison, requiring only a stable 160 MB $\Delta$GPU / 1860 MB peak GPU across all distortion levels (not plotted).