Mitigating Perceptual Judgment Bias in Multimodal LLM-as-a-Judge via Perceptual Perturbation and Reward Modeling

Researchers identify and address Perceptual Judgment Bias in multimodal large language models used as automated evaluators, where these models favor plausible narratives over visually accurate answers when text and images conflict. The team develops a training framework using perceptually perturbed datasets and reward modeling that improves MLLM judges' visual grounding and evaluation consistency across benchmarks.

Multimodal large language models have emerged as promising tools for automated evaluation tasks, yet this research exposes a fundamental vulnerability in their decision-making architecture. When presented with visual and textual information that contradicts each other, these models systematically prioritize narrative coherence over perceptual accuracy—a bias with serious implications for any evaluation system relying on genuine visual understanding.

The phenomenon identified here reflects a broader challenge in AI alignment: ensuring models behave according to their stated capabilities rather than exploiting easier shortcuts in their training data. This vulnerability becomes critical in applications where visual verification matters—from content moderation to scientific image analysis to quality assurance systems. The fact that models anchor to text rather than visual evidence suggests they haven't genuinely learned robust multimodal reasoning but instead developed sophisticated pattern-matching that privileges linguistic signals.

The researchers' solution leverages controlled perturbations to create counterfactual training data where perceptual errors are isolated and verifiable. By combining structured reward modeling with batch-ranking objectives, their framework forces models to ground evaluations in actual visual perception rather than textual plausibility. This approach sidesteps the need for exhaustive pairwise annotations while improving both ranking coherence and human-evaluation alignment.

The advancement carries implications beyond academic evaluation frameworks. As multimodal models proliferate in commercial applications—from autonomous systems to content curation—the ability to verify that models genuinely process visual information becomes economically important. Organizations deploying MLLM judges need confidence that these systems won't be fooled by clever text that contradicts visual reality. This work establishes practical methods for that verification and correction.

• →Multimodal LLM judges systematically prioritize textual narratives over visual evidence when they conflict, undermining evaluation reliability
• →Controlled visual perturbations reveal that models anchor to text rather than genuinely grounding decisions in visual perception
• →The Perceptually Perturbed Judgment Dataset enables scalable, verifiable supervision without extensive pairwise labeling
• →Combined GRPO-based reward and batch-ranking training improves perceptual fidelity and consistency across diverse benchmarks
• →This approach addresses a critical vulnerability relevant to any system relying on trustworthy multimodal evaluation and decision-making