Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Mapping User Trust in Vision Language Models: Research Landscape, Challenges, and Prospects

Agnese Chiatti, Sara Bernardini, Lara Shibelski Godoy Piccolo, Viola Schiaffonati, Matteo Matteucci

TL;DR

The paper addresses the challenge of trusting Vision Language Models (VLMs) by proposing a multidisciplinary framework to study trust dynamics in human–VLM interactions. It introduces a taxonomy extending the ABI trust model with Situated Cognition and Theory of Mind concepts, and applies this framework to a systematic review of 43 recent works and a pilot user study. The authors identify gaps, including limited attention to user-perceived Benevolence and cognitive Abilities, scarce direct trust-in-use studies, and a heavy emphasis on explicability and robustness. They provide preliminary, user-centered insights and concrete requirements for future VLM trust research, including the value of graph-based representations and ongoing trust monitoring. Collectively, the work lays groundwork for designing trustworthy VLM deployments in real-world, safety-critical contexts by guiding future multidisciplinary user studies and evaluation tools.

Abstract

The rapid adoption of Vision Language Models (VLMs), pre-trained on large image-text and video-text datasets, calls for protecting and informing users about when to trust these systems. This survey reviews studies on trust dynamics in user-VLM interactions, through a multi-disciplinary taxonomy encompassing different cognitive science capabilities, collaboration modes, and agent behaviours. Literature insights and findings from a workshop with prospective VLM users inform preliminary requirements for future VLM trust studies.

Mapping User Trust in Vision Language Models: Research Landscape, Challenges, and Prospects

TL;DR

The paper addresses the challenge of trusting Vision Language Models (VLMs) by proposing a multidisciplinary framework to study trust dynamics in human–VLM interactions. It introduces a taxonomy extending the ABI trust model with Situated Cognition and Theory of Mind concepts, and applies this framework to a systematic review of 43 recent works and a pilot user study. The authors identify gaps, including limited attention to user-perceived Benevolence and cognitive Abilities, scarce direct trust-in-use studies, and a heavy emphasis on explicability and robustness. They provide preliminary, user-centered insights and concrete requirements for future VLM trust research, including the value of graph-based representations and ongoing trust monitoring. Collectively, the work lays groundwork for designing trustworthy VLM deployments in real-world, safety-critical contexts by guiding future multidisciplinary user studies and evaluation tools.

Abstract

The rapid adoption of Vision Language Models (VLMs), pre-trained on large image-text and video-text datasets, calls for protecting and informing users about when to trust these systems. This survey reviews studies on trust dynamics in user-VLM interactions, through a multi-disciplinary taxonomy encompassing different cognitive science capabilities, collaboration modes, and agent behaviours. Literature insights and findings from a workshop with prospective VLM users inform preliminary requirements for future VLM trust studies.
Paper Structure (22 sections, 17 figures, 1 table)

This paper contains 22 sections, 17 figures, 1 table.

Table of Contents

  1. Vision Language Models (VLMs)
  2. Mapping User Trust in VLMs
  3. Search and selection criteria
  4. Proposed taxonomy
  5. Coverage study
  6. Methods and studies.
  7. User trust in VLMs: a case study
  8. Use-case scenario.
  9. Materials and methods
  10. Preparatory activities.
  11. Part 1: Collaborative game.
  12. Part 2: Mock-up evaluation.
  13. Participants
  14. Results
  15. Collaborative game.
  16. ...and 7 more sections

Figures (17)

  • Figure 1: Proposed taxonomy for modelling user trust in Vision Language Models. The size of leaf nodes is proportional to the count of surveyed papers for each category.
  • Figure 2: Question examples within each task, by increasing difficulty from basic perception and model building to advanced abstraction (learning to learn).
  • Figure 3: Structure of the workshop.
  • Figure 4: Design and Development workshop: players interact in pairs using different prompts to compare responses from the blindfolded player, the LLM, and the VLM.
  • Figure 5: Question structure in the Miro digital board.
  • ...and 12 more figures