Research map / Large language models / Language, safety and interpretability
MLLM-as-a-Judge: Assessing Multimodal LLM-as-a-Judge with Vision-Language Benchmark
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces MLLM-as-a-Judge, a benchmark of 4,414 image-instruction pairs with ~17k MLLM responses and human annotations to test how well MLLMs judge via scoring, pair comparison, and batch ranking.
Problem
Whether multimodal LLMs can act as reliable judges of multimodal responses and how closely their judgments align with human preferences is unclear, since traditional metrics fail to capture nuanced quality.
Approach
Builds a multimodal LLM-as-a-judge benchmark across three judging settings (scoring, pair comparison with ties, batch ranking) on 10-14 datasets, compares MLLM judgments to human annotations, and probes biases (position, length, egocentric), hallucination, and consistency, including vision-expert descriptions fed to text-only LLMs.
Key results
- GPT-4V achieves the highest Pearson similarity with human scores in Scoring Evaluation (0.490 average), while Gemini reaches only 0.304.
- In Pair Comparison with ties, GPT-4V averages 0.636 accuracy versus 0.509 for Gemini; GPT-4V reaches 79.3% human agreement in pair comparison.
- GPT-4V's batch ranking edit distance averages 0.361, the best among evaluated MLLMs, but all models remain far from human rankings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GPT-4V rather than Gemini as the judge | MLLM-as-a-Judge · Scoring Evaluation (average) | Pearson similarity: 0.304 -> 0.490 | Helped |
| Use GPT-4V rather than Gemini as the judge | MLLM-as-a-Judge · Pair Comparison with ties (average) | Pair similarity: 0.509 -> 0.636 | Helped |
| Use GPT-4V rather than Gemini as the judge | MLLM-as-a-Judge · Pair Comparison without ties (average) | Pair similarity: 0.615 -> 0.773 | Helped |
| Use GPT-4V rather than Gemini as the judge | MLLM-as-a-Judge · Batch Ranking (average) | Normalized Levenshtein distance: 0.432 -> 0.361 | Helped |
| Add 3-step CoT (ablation) | GPT-4V · MLLM-as-a-Judge selected subset | Pearson similarity: 0.557 -> 0.299 | Hurt |
| Add 3-step CoT (ablation) | Gemini · MLLM-as-a-Judge selected subset | Pearson similarity: 0.299 -> 0.144 | Hurt |
| Add 3-step CoT (ablation) | GPT-4V · Pair Comparison without ties | Pair similarity: 0.806 -> 0.728 | Hurt |
| Add 3-step CoT (ablation) | Gemini · Pair Comparison without ties | Pair similarity: 0.723 -> 0.377 | Hurt |
| Replace image input with a detailed vision-expert description (ablation) | GPT-4V · Scoring Evaluation | Pearson similarity: 0.299 -> 0.435 | Helped |
| Replace image input with a detailed vision-expert description (ablation) | GPT-4V · Pair Comparison without ties | Pair accuracy: 0.868 -> 0.878 | Helped |
| Replace image input with a detailed vision-expert description (ablation) | GPT-4V · Batch Ranking | Edit distance: 0.394 -> 0.400 | Hurt |
| Add extra reasoning steps before the usual Analyze-then-Judge process (ablation) | MLLM-as-a-Judge-Hard · Scoring Evaluation, Pair Comparison, and Batch Ranking | Hallucination frequency | Helped |
Surprising findings
- Multi-step (3-step) chain-of-thought does not improve alignment with humans and often degrades judging performance, especially for Gemini.
- Text-only LLMs given detailed GPT-4V image descriptions as vision experts can match or slightly outperform some MLLMs as judges.
Limitations
- Human annotation and MLLM judgments contain biases.
- Many models (CogVLM, LLaVA) fail to follow complex judging instructions, limiting comparisons.
- Gemini refusals and multilingual output issues, and GPT-4V refusals, complicate evaluation.
Experimental setup
Datasets: Conceptual Captions ChartQA InfographicVQA MathVista TextVQA WIT MS COCO DiffusionDB CC-3M VisIT-Bench Mind2Web AesBench
Benchmarks: MLLM-as-a-Judge
Metrics: Pearson correlation accuracy F1 recall normalized Levenshtein distance human agreement
Baselines: LLaVA-1.5-13b LLaVA-1.6-34b Gemini-Pro Qwen-VL-Max CogVLM GPT-4
Builds on: LLM-as-a-Judge Analyze-then-Judge Chain-of-Thought
Closest papers by meaning
- On Path to Multimodal Generalist: General-Level and General-Bench (ICML 2025 · similarity 0.93)
- JudgeLM: Fine-tuned Large Language Models are Scalable Judges (ICLR 2025 · similarity 0.92)
- SEED-Bench: Benchmarking Multimodal Large Language Models (CVPR 2024 · similarity 0.91)
- JudgeBench: A Benchmark for Evaluating LLM-Based Judges (ICLR 2025 · similarity 0.91)
- Bridging Human and LLM Judgments: Understanding and Narrowing the Gap (NeurIPS 2025 · similarity 0.90)
- Tuning LLM Judge Design Decisions for 1/1000 of the Cost (ICML 2025 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.