Research Summary: When the Judge Acts: Auditing VLM-Guided Image Selection on Culturally Situated Prompts

A research audit of Vision-Language Models (VLMs) acting as 'judges' for image selection revealed significant reliability and bias concerns. The audited VLM, despite its parameter count, demonstrated performance barely exceeding random selection and falling short of a basic similarity baseline. Key issues identified include a strong bias towards selecting the first image presented and high sensitivity to the order in which candidate images are displayed, indicating a lack of robust decision-making consistency.

Why it matters

This research highlights critical limitations in the reliability and consistency of current VLM technology when deployed as decision-makers in image selection processes. Organizations relying on or considering the use of VLMs for content moderation, selection, or ranking must account for inherent biases and order-dependency, which can lead to unpredictable outcomes and undermine user experience or operational integrity.

Key insights

• VLM judges tasked with image selection demonstrated performance only marginally better than random chance when evaluated against independent human ratings.
• The audited VLM performed worse than a simpler CLIP similarity baseline in its selection accuracy.
• A significant bias was observed, with the VLM judge selecting the first image presented in 49% of calls, substantially higher than the 28% expected by chance.
• The order of candidate images heavily influenced VLM decisions, with reordering changing the selected image on 60% of prompts.
• Agreement on image selection across different presentation orders was effectively zero for the evaluated VLM, highlighting severe inconsistency.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2610.01243