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

Source: arXiv - 2606.02578v1

Overview

Multimodal large language models (MLLMs) are increasingly being used as automated judges for tasks that involve both text and images—think content moderation, visual QA, or AI‑generated art scoring. This paper uncovers a systematic flaw the authors call Perceptual Judgment Bias: when visual evidence clashes with the narrative in the answer, the model tends to favor the plausible story rather than the actual visual content. By exposing and correcting this bias, the work paves the way for more trustworthy AI evaluators that can be safely deployed in production pipelines.

Key Contributions

• Identification of Perceptual Judgment Bias – a thorough analysis showing that current MLLM judges over‑rely on textual cues and ignore contradictory visual signals.
• Perceptually Perturbed Judgment Dataset (PPJD) – a curated collection of minimally edited, counterfactual responses that isolate pure perceptual errors, enabling clean supervision.
• Unified Training Framework – combines a structured GRPO (Generalized Rank‑Based Preference Optimization) reward with a batch‑ranking objective to enforce globally coherent rankings without needing exhaustive pairwise labels.
• Scalable Evaluation Protocol – extensive benchmarking across multiple MLLM‑as‑a‑Judge suites, demonstrating consistent gains in perceptual fidelity and alignment with human judgments.
• Open‑source Release – code, dataset, and trained reward models are publicly released, facilitating reproducibility and downstream adoption.

Methodology

1. Bias Diagnosis – The authors construct “visual‑conflict” test cases where the image clearly contradicts the answer’s claim (e.g., a red ball labeled as blue). They then measure how often the MLLM judge still assigns a high score to the misleading answer.
2. Dataset Construction (PPJD) – Starting from existing multimodal QA pairs, they apply tiny visual perturbations (color swaps, object insertions) and generate counterfactual textual responses that are identical except for the perceptual error. This isolates the visual component while keeping language constant.
3. Reward Modeling –
   • GRPO Reward: a structured reward that captures hierarchical preferences (e.g., “correct perception > plausible story”).
   • Batch‑Ranking Objective: during training, a batch of candidate answers is ranked jointly, encouraging the model to produce a globally consistent ordering rather than isolated pairwise decisions.
4. Training Loop – The multimodal judge is fine‑tuned with the combined loss (GRPO + batch ranking) on PPJD, while preserving its original language reasoning capabilities.

The approach is deliberately lightweight: it does not require annotating every possible answer pair, and the perturbations are automatically generated, making the pipeline scalable to large corpora.

Results & Findings

• Perceptual fidelity jumps by ~16 percentage points, meaning the judge now correctly penalizes answers that contradict the image.
• Ranking coherence improves dramatically, indicating that the model’s scores are internally consistent across a batch of candidates.
• Human evaluation shows the fine‑tuned judges are markedly closer to crowd‑sourced ratings, confirming that the bias mitigation translates to real‑world perception.
• The improvements hold across diverse domains (visual commonsense reasoning, meme captioning, AI‑generated art assessment), suggesting strong generalizability.

Practical Implications

• Content Moderation – Platforms can rely on MLLM judges to flag images that are misdescribed (e.g., deep‑fakes or misleading captions) with higher confidence.
• Automated Grading – Educational tools that score visual‑textual assignments (e.g., diagram explanations) will produce fairer grades because the model now respects the visual evidence.
• AI‑Generated Media Evaluation – Artists and product teams can use the refined judge to objectively compare generated images, ensuring that aesthetic scores are not inflated by clever but inaccurate descriptions.
• Reduced Human Oversight – Higher alignment with human judgments means fewer manual review cycles, cutting operational costs for large‑scale pipelines.
• Plug‑and‑Play Reward Model – The released reward model can be attached to any existing multimodal LLM (e.g., GPT‑4V, LLaVA) to instantly improve its judgment behavior without retraining the entire backbone.

Limitations & Future Work

• Scope of Perturbations – The PPJD focuses on low‑level visual changes (color, object presence). More complex semantic alterations (scene layout, abstract concepts) remain untested.
• Model Size Dependency – Gains are most pronounced on medium‑sized MLLMs; very large models (e.g., GPT‑4V) exhibit a smaller relative improvement, hinting at diminishing returns.
• Human Bias Transfer – The dataset inherits any biases present in the original human‑written answers, which could propagate into the reward model.
• Future Directions – Extending perturbation strategies to 3‑D or video data, incorporating explicit visual attention supervision, and exploring multi‑judge ensembles to further boost robustness.

Bottom line: By exposing and correcting the tendency of multimodal LLMs to favor plausible narratives over visual truth, this work delivers a practical, scalable recipe for building AI judges that are both perceptually grounded and reliable—an essential step toward trustworthy AI‑driven evaluation in real‑world systems.

Authors

• Seojeong Park
• Jiho Choi
• Junyong Kang
• Seonho Lee
• Jaeyo Shin
• Hyunjung Shim

Paper Information

• arXiv ID: 2606.02578v1
• Categories: cs.CV, cs.AI
• Published: June 1, 2026
• PDF: Download PDF