PaperlayerView exact revision on arXiv

Source-linked AI summary

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

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

arXiv:2606.02578v1cs.CVcs.AI

TL;DR

MLLM judges can favor linguistically plausible responses over visually correct ones, undermining reliable multimodal evaluation. This paper introduces perception-perturbed supervision and unified reward modeling, improving perceptual fidelity, ranking coherence, and alignment with human judgments across benchmarks.

  • Problem

    MLLM judges may prioritize linguistically plausible responses over visual correctness, creating unreliable evaluations when textual claims conflict with image evidence.

  • Method

    The paper constructs PPJD counterfactual response quadruplets and trains judges with structured GRPO rewards plus batch-ranking supervision for globally consistent ordering.

  • Results

    Across diverse benchmarks, the approach improves perceptual fidelity, ranking coherence, and human alignment, with batch-level metrics improving by up to 11% on Qwen3-VL-4B-Thinking and single-score prediction accuracy by 15% on Flex-Judge-7B.

  • Takeaways & Limitations

    The findings support scalable, perception-aware multimodal evaluation with globally consistent rankings and closer alignment to human judgments.

  • Takeaways & Limitations

    Residual biases and failure modes may persist in subjective, ambiguous, or out-of-distribution cases, requiring human review and domain-appropriate auditing for high-stakes use.

Abstract

from arXiv · show

Recent multimodal large language models have demonstrated strong reasoning ability, yet their reliability as automated evaluators remains limited by a critical weakness: when visual evidence conflicts with textual cues, MLLM judges tend to reward plausible narratives over perceptually correct answers. We identify and systematically analyze this phenomenon, which we term Perceptual Judgment Bias. Through controlled visual perturbations, existing multimodal judges frequently anchor on the response text instead of their own visual perception, leading to inconsistent and non-verifiable evaluations. To address this issue, we introduce the Perceptually Perturbed Judgment Dataset, which constructs minimally edited counterfactual responses that isolate perceptual errors and enable verifiable supervision. Building on this dataset, we develop a unified training framework that combines a structured GRPO-based reward with a batch-ranking objective, achieving coherent global ordering without explicit pairwise labels. Experiments across diverse MLLM-as-a-Judge benchmarks show that our approach substantially improves perceptual fidelity, ranking coherence, and alignment with human evaluation. Our results establish a scalable and generalizable pathway for training multimodal judges that are perceptually grounded, interpretable, and robust to visual-reasoning conflicts.

1. Introduction

The introduction identifies Perceptual Judgment Bias as a core reliability problem in multimodal LLM evaluators, which often favor plausible textual narratives over visual evidence. It presents Perception-Judge, trained with the PPJD dataset and structured objectives, as a scalable approach that improves perceptual grounding and evaluation consistency.

  • Motivation: MLLM response evaluation remains challenging despite progress in integrating textual and visual information.Human judgment is costly, time-consuming, and subject to inter-annotator inconsistency, motivating automated evaluators.
  • Problem: Perceptual Judgment Bias causes MLLM judges to produce linguistically plausible but visually ungrounded evaluations.The bias includes insufficient perceptual capability and response anchoring, in which judges rely on visual descriptions in response text rather than their own perception.
  • Problem: Baseline judges incur overall error rates exceeding 20%, showing that both insufficient perception and response anchoring degrade judgment performance.The introduction states that these error rates are quantified in Table 1.
  • Method: PPJD injects controlled visual inconsistencies into correct responses to create quadruplets that disentangle perceptual failures from reasoning errors.The dataset provides fine-grained, verifiable supervision for perception-aware alignment.
  • Results: 11% improvement on Qwen3-VL-4B-Thinking in batch-level metrics and 15% higher single-score prediction accuracy on Flex-Judge-7B demonstrate consistent empirical gains.The approach also matches leading proprietary evaluators in score-based settings and surpasses many in batch-level consistency.

2. Background & Preliminary

This section reviews multimodal LLM-as-a-Judge and visual reasoning, then introduces GRPO as a reinforcement-learning framework using group-normalized rewards and regularized policy updates. It formalizes candidate generation, relative-advantage computation, and the policy objective used for optimization.

  • MLLM-as-a-Judge & Visual Reasoning: LLM-as-a-Judge methods evaluate the quality of model-generated responses, with recent work incorporating reinforcement learning and verifiable rewards to improve scoring reliability and preference consistency.
  • Group Relative Policy Optimization (GRPO): Traditional PPO requires an explicit value network and can become unstable under sparse or noisy rewards, motivating GRPO as an alternative optimization approach.
  • Group Relative Policy Optimization (GRPO): Given prompt x from dataset D, the old policy πθold generates n candidate responses O = {o_i}^n_{i=1} for group-based reward optimization.
  • Group Relative Policy Optimization (GRPO): GRPO computes each response’s relative advantage by normalizing its reward against the group mean and standard deviation.The advantage is defined as Â_i = (R(o_i) − µ(R)) / σ(R).
  • Group Relative Policy Optimization (GRPO): The policy update increases the likelihood of higher-quality responses while constraining deviation from behavior and reference policies through PPO-style clipping and KL regularization.The probability ratio is r_i = πθ(o_i|x) / πθold(o_i|x), and β controls regularization.

3. Proposed Method

The proposed method formalizes multimodal judging, identifies Perceptual Judgment Bias as a failure to prioritize visual evidence over plausible response text, and constructs controlled perturbations to train more perceptually grounded evaluators. PPJD organizes correct, perception-perturbed, and fully degraded responses into graded supervision for GRPO-based alignment.

  • 3.1 MLLM-as-a-Judge Task: The MLLM-as-a-Judge task evaluates K responses to a textual instruction and visual input, producing scalar scores, pairwise preferences, or ranked candidate sequences.Judgments are intended to measure alignment with multimodal context and human evaluation.
  • Perceptual Judgment Bias: Perceptual Judgment Bias occurs when judges fail to penalize visually contradicted responses because they misperceive the image or anchor on linguistically plausible reasoning.The second failure mode arises even when the judge answers direct visual queries correctly but does not apply that perception during evaluation.
  • Perceptual Judgment Bias: Baseline judges incur overall error rates of 30.5% and 23.5%, with response-anchored errors comparable to or exceeding perceptual-recognition errors.This decomposition shows that the problem includes failure to connect recovered visual evidence with evaluative reasoning, not only image-recognition mistakes.
  • Perception-Guided Training: The training framework uses quadruplets in GRPO to impose the graded judgment order rc ≻ rrp ≻ rrp+r without requiring a value function.This ordering separates isolated perceptual errors from responses degraded in both perception and reasoning, while GRPO optimizes relative advantages within grouped responses.
  • Perceptually Perturbed Judgment Dataset: PPJD starts from correct responses and generates a reasoning-preserved perceptual perturbation rrp plus a response with both perceptual and reasoning inconsistencies, rrp+r.Perceptual perturbations alter grounded attributes such as color, count, or spatial relations while preserving the original response structure, forming quadruplets with the input and correct response.

4. Experiments

Experiments show that Perception-Judge improves perceptual fidelity, pairwise performance, and batch-level ranking consistency over strong multimodal baselines. Gains are attributed to the compact PPJD dataset and globally consistent batch-ranking supervision.

  • Evaluation: The evaluation measures Pearson correlation for single-score grading, accuracy, F1 score, and recall for pairwise comparisons, and normalized Levenshtein distance for batch rankings.These metrics quantify agreement between model judgments and human annotations across evaluation settings.
  • Quantitative Results: 4% improvement in pairwise evaluations and 11% improvement in batch-level evaluations are reported over Qwen3-VL-4B-Thinking.The method improves across most benchmark settings despite using no explicit score labels during training.
  • Quantitative Results: Perception-Judge is comparable to GPT-4o in single-score prediction and outperforms most proprietary LLMs in batch-level evaluation.The batch-level advantage reflects stronger global consistency across multi-response comparisons.
  • Data Efficiency: A compact, evaluation-disjoint 3k PPJD dataset achieves competitive performance with substantially less supervision than the 113k-example LLaVA-Critic corpus.LLaVA-Critic attains higher pairwise scores, but the comparison highlights PPJD’s data efficiency.
  • Ablation Study: Training on PPJD instead of MMPR-v1.2 yields consistent improvements across all metrics, indicating that perceptual corrections make multimodal evaluation more reliable.PPJD encodes corrected–perceptual and perceptual–perturbed comparisons through preference pairs.
  • Ablation Study: Batch-ranking reward outperforms pairwise reward on score-based and pairwise metrics while enforcing globally consistent candidate ordering without explicit pairwise or score labels.The comparison uses PPJD with pairwise reward versus PPJD with batch ranking reward.

5. Conclusion

The work identifies Perceptual Judgment Bias as a failure mode in multimodal LLM judges and proposes perception-guided training to improve perceptual reliability and human-evaluation consistency.

  • 5. Conclusion: Perceptual Judgment Bias causes multimodal LLM judges to prioritize linguistic plausibility over visual correctness, producing systematic divergence from human evaluation.The paper frames this as a fundamental failure mode affecting perceptual reliability.
  • 5. Conclusion: The proposed perception-guided training framework is built upon the Perceptually Perturbed Judgment Dataset (PPJD).PPJD provides the dataset foundation for addressing the identified limitation.
  • 5. Conclusion: The framework combines unified GRPO-based optimization with a structured batch reward mechanism.This unified optimization design is presented as the training approach for mitigating perceptual judgment bias.

Impact Statement

The framework aims to improve the reliability, fairness, transparency, and human alignment of automated multimodal evaluators by mitigating perceptual judgment bias. However, residual biases and failure modes remain, especially in subjective, ambiguous, or out-of-distribution cases, creating risks if automated judges replace human oversight or are treated as definitive authorities.

  • Benefits: The framework targets more reliable and fairer multimodal evaluators that are transparent, visually grounded, and aligned with human judgment.This motivation is especially relevant as MLLMs enter settings requiring accurate visual understanding and faithful evaluation.
  • Benefits: The approach can reduce evaluation costs and reliance on extensive human annotation.
  • Risks: Residual biases and failure modes may persist in subjective, ambiguous, or out-of-distribution cases, particularly when automated judges replace human oversight.These risks could be amplified if such systems are treated as definitive authorities.

A. Limitations & Future Work … C.1. Code & Reproduction

The paper identifies limitations in judge performance, supervision dependence, subjective ambiguity, and computational demands while situating its method among multimodal judging and visual-perception research. It also documents a standardized experimental setup and releases a PyTorch/verl implementation for reproduction.

  • A. Limitations & Future Work: Judge performance still trails expert human annotators in some challenging scenarios, motivating improved reasoning and more diverse preference signals.The training method also requires more computation than a truncated comparison described in the passage.
  • A. Limitations & Future Work: Proprietary supervisory models limit reproducibility and applicability, motivating self-distillation from capable open-source models such as Qwen.The passage also identifies extending the method to larger model scales as future work, but the sentence is truncated.
  • A. Limitations & Future Work: Subjective multimodal evaluation remains difficult because subtle interpretation ambiguities and residual reasoning noise can persist after filtering.The proposed procedure reduces variability but does not eliminate these challenges.
  • B. Related Work: MLLM-as-a-Judge extends language-model evaluation to vision–language responses, but existing approaches mainly use supervised fine-tuning and exhibit visual perception judgment bias.The bias favors linguistic plausibility over perceptual correctness, which motivates the paper’s approach.
  • B. Related Work: Prior visual-perception research attributes MLLM visual-reasoning difficulties partly to under-attending visual information when textual sequences dominate visual tokens.The passage surveys studies reporting limited visual perception capabilities and identifies this attention imbalance as one explanation.
  • B. Related Work: Complementary methods improve visual reasoning by using corrupted-image negatives, augmentation discrimination, or perception tokens.PAPO, MiCo, and LLaVA-AURORA represent these respective strategies.
  • C. Experimental Details: The experiments, including GRPO-based fine-tuning and multimodal-judgment benchmark evaluation, use the same standardized setup unless otherwise specified.The experimental-details passage introduces additional information intended to support reproducibility and clarify computational configuration.
  • C.1. Code & Reproduction: The training and evaluation pipeline is implemented in PyTorch on the verl framework, with code released in the perception-judge GitHub repository.verl supports reinforcement learning with verifiable rewards.

C.2. Implementation Details & Hyperparameters … D.1. PPJD Statistics

The paper specifies training, evaluation, bias-mitigation, and prompting procedures for Perception-Judge, alongside baseline descriptions and PPJD dataset statistics. Evaluation spans single-score grading, pairwise comparison, and batch-level ranking under standardized protocols.

  • C.2. Implementation Details & Hyperparameters: Perception-Judge trains Flex-Judge-VL and Qwen3-VL-Thinking with GRPO, five rollouts per prompt, one PPJD epoch, and global batch size 512.Learning rates are 2.5 × 10−7 for Flex-Judge-VL and 7.5 × 10−7 for Qwen3-VL-Thinking.
  • C.2. Implementation Details & Hyperparameters: Training uses 8 NVIDIA H100 GPUs for Perception-Judge-Flex and 8 NVIDIA H200 GPUs for Perception-Judge-Qwen3, while inference uses the same configuration without tuning.Evaluation follows the standardized MLLM-as-a-Judge benchmark protocol.
  • C.3.1. EVALUATION SETTINGS: The evaluation protocol measures alignment with human preferences through single-score grading, pairwise comparison, and batch-level ranking.Single-score grading assigns scalar scores, pairwise comparison predicts preferences, and batch ranking compares predicted permutations with human rankings.
  • C.3.2. EVALUATION PROMPTS: The shared prompt templates require structured answers within <answer></answer> tags, with batch ranking explicitly disallowing identical scores across assistants.The templates cover single-score, pairwise, and batch-ranking evaluations.
  • C.3.2. EVALUATION PROMPTS: Prompt outputs are deterministically post-processed by mapping single-score results from 1 to 10 onto 1 to 5, deriving pairwise preferences, and ordering batch rankings.Pairwise ties are assigned when scores fall within the same range.
  • C.3.3. BIAS MITIGATION: Bias mitigation uses randomized response ordering for pairwise and batch settings and tie handling to avoid positional and forced-choice artifacts.When ties are unavailable, the model makes a forced random choice between candidates.
  • C.4. Baselines: Baselines include LLaVA-Critic, Qwen2.5-VL, Flex-Judge-VL, Prometheus-V, and Qwen3-VL, spanning judge-specialized, general-purpose, and zero-shot multimodal evaluators.The baselines differ in training data, model scale, evaluation specialization, and multimodal capabilities.
  • D.1. PPJD Statistics: PPJD contains 3K instances across visual question answering, multi-step reasoning, mathematical and geometric queries, document information extraction, and image quality assessment.These categories provide complementary visual signals and response patterns for multimodal judgment tasks.

D.2. Detailed Dataset Generation Pipeline … F.3. Pair Evaluation on F1 and Recall

The paper builds PPJD from absolutely validated MMPR answers using controlled perceptual and joint perceptual-reasoning perturbations, then evaluates Perception-Judge through reward dynamics, scaling, human-aligned scoring, and pairwise preference identification.

  • D.2. Detailed Dataset Generation Pipeline: PPJD represents each example as (x, rc, rrp, rrp+r), pairing a validated correct answer with perceptual and joint perceptual-reasoning rejected candidates.The dataset begins from MMPR and applies controlled perturbations to perceptual and reasoning attributes.
  • D.2. Detailed Dataset Generation Pipeline: MMPR preferred answers become rc only after matching the original dataset’s ground-truth label; mismatches are discarded.This filtering makes rc absolutely correct rather than merely relatively preferred.
  • D.2. Detailed Dataset Generation Pipeline: Perceptual rejected candidates minimally alter visually validated attributes while preserving answer structure, whereas joint candidates also introduce reasoning errors dependent on those altered attributes.Accepted candidates must remain internally coherent and diverge from ground truth.
  • E. Reward Analysis of Perception-Judge: Both Perception-Judge-Flex-7B and Perception-Judge-Qwen3-4B rapidly reach near-perfect RFormat, showing that output-format constraints are readily learnable.The analysis tracks RFormat and RBatch on training and validation splits of PPJD.
  • F.1. Generalization to a larger model scale: Scale-up studies report consistent performance gains for Flex-Judge-VL-32B and Qwen3-VL-8B-Thinking, supporting generalization across model sizes and architectures.These models extend experiments beyond the 7B and 4B configurations used in the main paper.
  • F.2. Score Evaluation on Spearman: Perception-Judge achieves substantially higher Spearman correlation with human judgments than baseline multimodal judges, indicating stronger monotonic agreement across perceptual tasks.The improvement is attributed to suppressing visually inconsistent responses and maintaining coherent scoring.
  • F.3. Pair Evaluation on F1 and Recall: Perception-Judge surpasses Qwen2.5-VL-7B and Flex-Judge-VL-7B in both recall and F1 when identifying human-preferred responses.The result indicates greater sensitivity to visually grounded correctness and reduced susceptibility to fluent but perceptually incorrect outputs.

F.4. Batch Evaluation on Edit Distance · G. Additional Qualitative Results

Batch-level evaluation finds that Perception-Judge produces rankings more consistent with human preference orderings, with lower edit distance and higher Kendall’s tau and NDCG. The supplied passages provide no substantive content for the qualitative-results section.

  • F.4. Batch Evaluation on Edit Distance: Perception-Judge achieves lower edit distance in batch-level ranking evaluation.This indicates more stable global ranking structures relative to the evaluated alternatives.
  • F.4. Batch Evaluation on Edit Distance: Perception-Judge achieves higher Kendall’s tau in batch-level ranking evaluation.The metric contributes to evidence of stronger agreement with human preference orderings.
  • F.4. Batch Evaluation on Edit Distance: Perception-Judge achieves higher NDCG in batch-level ranking evaluation.The result supports more stable ranking structures that better reflect perceptual fidelity.
  • F.4. Batch Evaluation on Edit Distance: Batch-level ranking metrics quantify global consistency with human preference orderings.The reported metrics are lower edit distance, Kendall’s tau, and NDCG.
  • F.4. Batch Evaluation on Edit Distance: The results confirm benefits from perceptual grounding in large-scale evaluation scenarios.The conclusion follows from the improved batch-level ranking metrics.
  • F.4. Batch Evaluation on Edit Distance: Each judgment is sampled three times and averaged to reduce variance.This evaluation procedure applies to the reported judgments.
  • F.4. Batch Evaluation on Edit Distance: Table S3 presents comprehensive metric summaries for complementary evaluation measures.The listed measures include Spearman correlation, F1 score, Recall, Kendall’s tau, and NDCG.

G.1. MLLM-as-a-Judge Qualitative Results · H. Additional Ablation Study

Qualitative examples show that Perception-Judge corrects baseline judges’ perceptual failures by prioritizing visual evidence over fluent but incorrect response narratives. Across cases, it more reliably detects mismatches and reduces cognitive biases in MLLM-as-a-Judge evaluation.

  • G.1. MLLM-as-a-Judge Qualitative Results: Baseline judges, including Qwen2.5-VL-7B, Flex-Judge-VL-7B, and Qwen3-VL-4B-Thinking, exhibit inaccurate visual interpretation and response-anchored evaluation.These are the two recurring error modes highlighted in the qualitative examples.
  • G.1. MLLM-as-a-Judge Qualitative Results: Perception-Judge corrects systematic perceptual failures observed in baseline multimodal judges across additional qualitative examples.The evidence is presented in Figures S9, S12, S13, and S14.
  • G.1. MLLM-as-a-Judge Qualitative Results: Baseline judges often reward logically fluent responses that are perceptually incorrect.The cases demonstrate that logical fluency can receive high scores despite conflict with the visual input.
  • G.1. MLLM-as-a-Judge Qualitative Results: Perception-Judge identifies discrepancies between visual inputs and candidate responses, penalizing misaligned descriptions and favoring perceptually grounded reasoning.This behavior is shown across varied qualitative cases.
  • G.1. MLLM-as-a-Judge Qualitative Results: The qualitative examples reinforce that Perception-Judge aligns evaluation more reliably with the underlying visual evidence.This improvement is presented as a consistent pattern across the examples.
  • G.1. MLLM-as-a-Judge Qualitative Results: Perception-Judge mitigates cognitive biases prevalent in existing MLLM-as-a-Judge systems.The conclusion follows collectively from the qualitative examples.

H.1. Detailed Results of Ablation Study · H.2. Ablation studies on KL penalty · I. Failure Cases

The ablations show that PPJD improves perceptual judgment over MMPR, while batch reward provides more globally consistent supervision than pair reward and a KL coefficient of 1e-2 yields peak Flex-Judge-7B performance. Remaining failures arise from subtle or ambiguous perception, implicit 3D and spatial understanding, and complex multi-step reasoning.

  • H.1. Detailed Results of Ablation Study: PPJD consistently improves over MMPR across evaluated benchmarks, with especially pronounced gains on tasks requiring precise perceptual grounding.The results attribute these gains to PPJD’s perceptually aligned data, which better supports robust judgment behavior.
  • H.1. Detailed Results of Ablation Study: Batch reward provides more stable and globally consistent supervision than pair reward across score-based, pairwise, and ranking evaluations.Pair reward supplies a minimal local comparison signal, whereas batch reward introduces structured ranking information.
  • H.1. Detailed Results of Ablation Study: Pair reward trains local perceptual preferences by comparing responses with increasing degradation rather than optimizing absolute scores.The target preferences are rc ≻ rrp and rrp ≻ rrp+r, with pairs sampled equally from the two adjacent comparisons.
  • H.2. Ablation studies on KL penalty: The KL coefficient regularizes GRPO by limiting deviation from the reference policy and balancing exploration with stability.The ablation on Flex-Judge-7B identifies 1e-2 as the coefficient associated with peak performance.
  • I. Failure Cases: The judge can fail on extremely fine-grained or atypical perceptual recognition when visual cues are subtle or ambiguous.These failures are more pronounced for implicit three-dimensional structure and complex spatial relationships that current multimodal backbones do not fully interpret.
  • I. Failure Cases: Complex multi-step reasoning can still produce incorrect judgments when visual perception alone is insufficient.These cases motivate stronger base-model reasoning capabilities for fully reliable judgment.
Loading 2606.02578v1…