VLM-UQBench: A Benchmark for Modality-Specific and Cross-Modality Uncertainties in Vision Language Models

Technical Summary: VLM-UQBench

VLM-UQBench: A Benchmark for Modality-Specific and Cross-Modality Uncertainties in Vision Language Models

1. Problem Statement

As Vision-Language Models (VLMs) are increasingly deployed in high-stakes domains (e.g., medical analysis, robotics, assistive technology), the ability to quantify uncertainty (UQ) is critical for safety. Current UQ research primarily relies on the classical epistemic-aleatoric dichotomy (distinguishing between model uncertainty and data uncertainty). However, this binary view is insufficient for VLMs because uncertainty can originate from three distinct, interacting sources:

1. Visual Modality: Low-quality, blurry, or occluded images.
2. Textual Modality: Ambiguous, invalid, or subjective queries.
3. Cross-Modal Interaction: Misalignment or insufficient evidence linking the text to the image.

Existing benchmarks fail to provide the modality-aware, instance-wise, and contrastive structure necessary to determine where uncertainty arises, which is essential for deciding whether a system should retake an image, rephrase a question, or defer to a human.

2. Methodology

The authors introduce VLM-UQBench, a comprehensive benchmark designed to probe modality-specific uncertainty. The methodology consists of four pillars:

• Annotated Modality Subsets: Using the VizWiz dataset, the authors curated four subsets (Clean, Image, Text, and Cross-modal) through a pipeline of human annotation and expert PhD-level curation. This provides ground-truth labels for the source of uncertainty.
• Grounded Cross-Modality Subset: They integrate data from VQ-FocusAmbiguity, using segmentation-based grounding to provide explicit supervision for uncertainty arising from text-image alignment.
• Hallucination-Focused Subsets: Using the CLEVR dataset and its scene graphs, they constructed controlled sets to evaluate how uncertainty leads to specific hallucination types: attribute, existence, counting, and relation.
• Scalable Perturbation Pipeline: To allow the benchmark to scale to any dataset, they developed a pipeline that injects controllable uncertainty via:
  • Visual edits: Blur, noise, cutout, pixelation, etc.
  • Textual edits: LLM-based paraphrasing (subjective/invalid), typos, and word shuffling.
  • Cross-modal edits: CLIP-based attention masking and prompt-based ambiguity injection.

Evaluation Metrics:
The paper introduces two novel metrics to move beyond simple scalar confidence:

1. Uncertainty Reflection Rate (URR): Measures how consistently a UQ method increases its uncertainty score when a specific modality is perturbed.
2. Hallucination Consistency Coefficient (HCC): A point-biserial correlation metric that evaluates whether changes in uncertainty scores actually track the occurrence of hallucinations.

3. Key Contributions

• A New Benchmark Standard: The first benchmark to provide modality-specific uncertainty labels (image, text, cross-modal) for free-form VQA.
• Automated Perturbation Framework: A scalable pipeline that enables fine-grained, instance-wise contrastive evaluation without requiring manual labels for every variant.
• New Evaluation Metrics: The introduction of URR and HCC to specifically measure modality sensitivity and hallucination-risk correlation.
• Systematic Empirical Analysis: A large-scale study of nine UQ methods across four VLMs (including open-source models like LLaVA/Qwen and closed-source GPT-4o-mini).

4. Key Results

• Modality Specialization & Model Dependence: UQ methods are not "one size fits all." For example, Semantic Entropy excels at textual uncertainty, but visual uncertainty remains a major failure point for most models. Furthermore, the best UQ method changes depending on the underlying VLM architecture.
• Weak Hallucination Signals: A critical finding is the disconnect between UQ scores and hallucination risk. While modality-specific uncertainty (especially visual) significantly drives hallucinations, current UQ methods show weak and inconsistent correlation (low HCC) with these failures.
• Ambiguity Detection Gap: While UQ methods can rival reasoning-based (Chain-of-Thought) approaches in detecting "overt" group-level ambiguity, they largely fail to detect the "subtle" instance-level ambiguity introduced by the perturbation pipeline.

5. Significance

This work shifts the paradigm of VLM uncertainty quantification from "how uncertain is the model?" to "where is the uncertainty coming from?" By exposing the inability of current UQ methods to reliably detect visual-driven hallucinations and subtle cross-modal misalignments, the paper highlights a significant gap in the reliability of modern VLMs. VLM-UQBench provides the necessary infrastructure for the community to develop the next generation of modality-aware, source-aware uncertainty quantification techniques required for safe real-world deployment.