Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Seeing is Believing, but How Much? A Comprehensive Analysis of Verbalized Calibration in Vision-Language Models

Weihao Xuan, Qingcheng Zeng, Heli Qi, Junjue Wang, Naoto Yokoya

TL;DR

This work systematically evaluates how vision-language models verbalize uncertainty across multiple modalities and task settings, revealing widespread miscalibration and a notable modality gap. It shows that vision-centered, reasoning-focused models tend to calibrate better than text-centric or instruction-tuned counterparts, and introduces Visual Confidence-Aware Prompting (VCAP) to improve calibration by decoupling visual perception from task reasoning. The study underscores the importance of modality alignment and faithful uncertainty representations for reliable multimodal systems, and provides a blueprint for calibration-aware prompting in practical applications. The findings advocate integrating modality-specific reasoning signals into prompting and evaluation to enhance trustworthiness in multimodal AI systems.

Abstract

Uncertainty quantification is essential for assessing the reliability and trustworthiness of modern AI systems. Among existing approaches, verbalized uncertainty, where models express their confidence through natural language, has emerged as a lightweight and interpretable solution in large language models (LLMs). However, its effectiveness in vision-language models (VLMs) remains insufficiently studied. In this work, we conduct a comprehensive evaluation of verbalized confidence in VLMs, spanning three model categories, four task domains, and three evaluation scenarios. Our results show that current VLMs often display notable miscalibration across diverse tasks and settings. Notably, visual reasoning models (i.e., thinking with images) consistently exhibit better calibration, suggesting that modality-specific reasoning is critical for reliable uncertainty estimation. To further address calibration challenges, we introduce Visual Confidence-Aware Prompting, a two-stage prompting strategy that improves confidence alignment in multimodal settings. Overall, our study highlights the inherent miscalibration in VLMs across modalities. More broadly, our findings underscore the fundamental importance of modality alignment and model faithfulness in advancing reliable multimodal systems.

Seeing is Believing, but How Much? A Comprehensive Analysis of Verbalized Calibration in Vision-Language Models

TL;DR

This work systematically evaluates how vision-language models verbalize uncertainty across multiple modalities and task settings, revealing widespread miscalibration and a notable modality gap. It shows that vision-centered, reasoning-focused models tend to calibrate better than text-centric or instruction-tuned counterparts, and introduces Visual Confidence-Aware Prompting (VCAP) to improve calibration by decoupling visual perception from task reasoning. The study underscores the importance of modality alignment and faithful uncertainty representations for reliable multimodal systems, and provides a blueprint for calibration-aware prompting in practical applications. The findings advocate integrating modality-specific reasoning signals into prompting and evaluation to enhance trustworthiness in multimodal AI systems.

Abstract

Uncertainty quantification is essential for assessing the reliability and trustworthiness of modern AI systems. Among existing approaches, verbalized uncertainty, where models express their confidence through natural language, has emerged as a lightweight and interpretable solution in large language models (LLMs). However, its effectiveness in vision-language models (VLMs) remains insufficiently studied. In this work, we conduct a comprehensive evaluation of verbalized confidence in VLMs, spanning three model categories, four task domains, and three evaluation scenarios. Our results show that current VLMs often display notable miscalibration across diverse tasks and settings. Notably, visual reasoning models (i.e., thinking with images) consistently exhibit better calibration, suggesting that modality-specific reasoning is critical for reliable uncertainty estimation. To further address calibration challenges, we introduce Visual Confidence-Aware Prompting, a two-stage prompting strategy that improves confidence alignment in multimodal settings. Overall, our study highlights the inherent miscalibration in VLMs across modalities. More broadly, our findings underscore the fundamental importance of modality alignment and model faithfulness in advancing reliable multimodal systems.

Paper Structure

This paper contains 42 sections, 1 equation, 4 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Uncertainty Quantification and Calibration in L(V)LMs
  4. Modality Misalignment in Multimodal LLMs
  5. Experimental Setup
  6. Evaluation Configurations
  7. General Evaluation
  8. Embedded Instruction Evaluation
  9. Semantically Aligned Modalities Evaluation
  10. Models
  11. Evaluation Metrics
  12. Results
  13. General Evaluation
  14. Embedded Instruction Evaluation
  15. Semantically Aligned Evaluation
  16. ...and 27 more sections

Figures (4)

  • Figure 1: The illustration of our three types of evaluations: general, embedded instruction, and semantically aligned evaluation. These configurations test VLMs' calibration across different input modalities and instruction formats.
  • Figure 2: Calibration curve on testmini set of MathVision.
  • Figure 3: Model performance comparison across and calibration. The upper right indicates better overall performance.
  • Figure 4: The illustration of our Visual Confidence-Aware Prompting (VCAP).