Title: Enhancing Multimodal LLM Judges: A Strategy Using Perceptual Perturbation and Reward Modeling to Counter Judgment Bias

Abstract

While multimodal large language models (MLLMs) have exhibited impressive reasoning capabilities, their effectiveness as automated evaluators is hindered by a significant flaw: when visual data contradicts textual information, MLLM judges often favor coherent narratives rather than perceptually accurate answers. This study identifies and rigorously examines this issue, which we define as Perceptual Judgment Bias. Our analysis reveals that through controlled visual perturbations, current multimodal judges tend to rely on response text rather than their own visual input, resulting in evaluations that are both inconsistent and unverified. To resolve this challenge, we present the Perceptually Perturbed Judgment Dataset, a resource designed to isolate perceptual errors by generating minimally altered counterfactual responses, thereby facilitating verifiable supervision. Leveraging this dataset, we propose a unified training framework that integrates a structured GRPO-based reward mechanism with a batch-ranking objective. This approach ensures consistent global ordering without the need for explicit pairwise labels. Our experimental results across various MLLM-as-a-Judge benchmarks demonstrate that this method significantly enhances perceptual fidelity, ranking coherence, and alignment with human judgment. These findings outline a scalable and generalizable route for developing multimodal judges that are robust against visual-reasoning conflicts, interpretable, and grounded in perception.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC