Fooling the LVLM Judges: Visual Biases in LVLM-Based Evaluation

2025-05-21

AI Summary

This work identifies a systematic visual-induction bias in large vision-language models (LVLMs) when used as image-text alignment evaluators: adversarial visual perturbations significantly inflate alignment scores, leading to distorted assessments. To address this, the authors formally define the bias and introduce FRAME—a fine-grained, multi-domain meta-evaluation benchmark. Using adversarial image generation, score distribution analysis, and pairwise evaluation protocols, they empirically demonstrate that the bias exhibits cumulative effects and remains unmitigated by mainstream prompt-engineering strategies. Experiments across all tested LVLM evaluators and domains reveal consistent, statistically significant score inflation. This study exposes critical fragility in current LVLM evaluation paradigms, providing a precise problem formulation, a standardized evaluation toolkit, and empirical evidence essential for developing robust image-text alignment scorers.

Technology Category

• Computer Vision: Adversarial Attacks & Robustness
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Language Grounding & Multi-modal NLP

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Recently, large vision-language models (LVLMs) have emerged as the preferred tools for judging text-image alignment, yet their robustness along the visual modality remains underexplored. This work is the first study to address a key research question: Can adversarial visual manipulations systematically fool LVLM judges into assigning unfairly inflated scores? We define potential image induced biases within the context of T2I evaluation and examine how these biases affect the evaluations of LVLM judges. Moreover, we introduce a novel, fine-grained, multi-domain meta-evaluation benchmark named FRAME, which is deliberately constructed to exhibit diverse score distributions. By introducing the defined biases into the benchmark, we reveal that all tested LVLM judges exhibit vulnerability across all domains, consistently inflating scores for manipulated images. Further analysis reveals that combining multiple biases amplifies their effects, and pairwise evaluations are similarly susceptible. Moreover, we observe that visual biases persist under prompt-based mitigation strategies, highlighting the vulnerability of current LVLM evaluation systems and underscoring the urgent need for more robust LVLM judges.

Problem

• Investigates visual biases in LVLM-based text-image alignment evaluation
• Examines adversarial image manipulations inflating LVLM judge scores
• Assesses vulnerability of LVLM judges across multi-domain benchmark

Innovation

• Adversarial visual manipulations inflate LVLM scores
• FRAME benchmark tests multi-domain visual biases
• Combined biases amplify LVLM judge vulnerability