AlignMMBench: Evaluating Chinese Multimodal Alignment in Large Vision-Language Models

Yuhang Wu*1, Wenmeng Yu*2, Yean Cheng3, Yan Wang2,
Xiaohan Zhang2, Jiazheng Xu1, Ming Ding2, Yuxiao Dong1
1Tsinghua University, 2Zhipu AI, 3Peking University,
*Indicates Equal Contribution
arXiv Code

📊

 Dataset

🏆

 Leaderboard

Abstract

Evaluating the alignment capabilities of large Vision-Language Models (VLMs) is essential for determining their effectiveness as helpful assistants. However, existing benchmarks primarily focus on basic abilities using nonverbal methods, such as yes-no and multiple-choice questions. In this paper, we address this gap by introducing AlignMMBench, a comprehensive alignment benchmark specifically designed for emerging Chinese VLMs. This benchmark is meticulously curated from real-world scenarios and Chinese Internet sources, encompassing thirteen specific tasks across three categories, and includes both single-turn and multi-turn dialogue scenarios. Incorporating a prompt rewrite strategy, AlignMMBench encompasses 1,054 images and 4,978 question-answer pairs. To facilitate the evaluation pipeline, we propose CritiqueVLM, a rule-calibrated evaluator that exceeds GPT-4's evaluation ability. Finally, we report the performance of representative VLMs on AlignMMBench, offering insights into the capabilities and limitations of different VLM architectures. All evaluation codes and data are available on https://alignmmbench.github.io/.

Leaderboard

Evaluation results on AlignMMBench. The "Ref." column indicates the relative ranking of these models on opencompass, dominated by primarily English benchmarks.

Models Size Ref. Avg Perception \& Understanding. Reasoning \& Analysis Context Align.
Des. Rec. Cou. OCR. Mem. Kno. Rea. Cha. Pro. Com. Wri. Coh. Inc.
Qwen2-VL 72B 1 6.51 7.39 6.64 6.64 7.60 7.09 6.32 4.00 7.16 5.89 6.57 7.72 6.37 5.26 1.54
Claude - 4 6.51 7.68 6.89 6.79 7.02 7.10 6.28 4.06 7.11 5.20 5.92 7.98 7.02 5.52 1.45
GPT-4o - 2 6.41 7.75 6.41 5.20 7.17 7.28 6.16 4.44 7.23 5.81 7.19 7.85 6.41 4.43 1.18
CogVLM2 19B 8 5.81 7.20 6.12 5.75 7.21 6.07 5.69 3.43 5.92 4.37 5.65 7.34 6.33 4.43 1.49
InternVL-Chat 26B 5 5.62 7.12 6.00 5.51 6.63 4.99 5.08 3.35 5.98 3.98 6.33 7.26 6.31 4.48 1.12
InternVL2 76B 3 5.57 6.95 5.11 5.81 7.37 5.96 3.61 3.83 6.48 4.66 6.05 6.05 6.30 4.23 0.93
MiniCPM 8B 6 5.42 7.18 5.37 5.46 6.23 4.46 5.35 3.34 4.83 3.69 5.99 7.35 6.25 4.97 1.09
Qwen-VL-Chat 9B 12 5.13 6.43 5.87 5.40 4.80 5.11 5.58 2.98 4.10 3.12 5.51 7.19 6.07 4.50 1.01
InternLM-XC2-VL 7B 7 4.97 6.34 4.70 5.28 5.06 4.69 5.03 3.08 4.49 3.29 5.00 7.21 5.92 4.56 0.88
DeepSeek-VL 7B 11 4.70 6.53 5.52 5.10 3.98 3.87 4.19 2.50 3.96 2.58 5.46 7.15 5.83 4.47 1.02
Monkey-Chat 9B 10 4.70 6.04 4.88 5.57 4.66 4.18 4.96 3.01 4.00 2.61 4.87 6.29 6.15 3.96 0.96
ShareGPT4V 13B 14 4.39 5.93 4.61 5.16 3.77 4.04 4.58 2.45 3.73 2.19 5.05 6.39 5.36 3.79 1.08
LLava-v1.5 13B 15 4.31 6.02 4.56 4.46 3.85 3.69 4.72 2.46 3.69 2.10 4.75 6.21 5.60 3.96 1.05
Yi-VL 34B 13 4.25 4.79 4.78 5.19 3.33 3.58 4.47 2.42 3.25 2.08 4.72 6.61 5.87 4.13 1.20
Phi-3-Vision 4B 9 4.08 4.48 3.53 4.75 4.10 3.48 3.16 2.56 4.40 2.85 4.34 5.51 5.85 4.07 0.86
InstructBLIP 9B 16 3.31 4.11 4.61 4.11 2.77 3.05 2.92 1.76 2.58 1.12 3.36 3.17 5.42 4.02 1.09
GPT-4o without image - - 2.13 1.11 1.57 1.22 1.73 1.53 1.17 1.29 2.88 1.14 1.99 3.50 5.14 3.41 -

Examples

Overall Framework

MY ALT TEXT

AlignMMBench: Images are collected from real-world requests and Chinese websites. Following manual taxonomy and LLM-based question expansion, annotators generate reference answers within task-specific rules. CritiqueVLM: Rule-based scoring criteria are carefully designed, and then we train the CritiqueVLM on human-annotated corpus.

Statistics

MY ALT TEXT

(Left) Categories and examples of AlignMMBench. The chart on the left displays the categories of AlignMMBench, encompassing three main categories and thirteen specific tasks. The numbers below each category or task indicate the total number of images included.

(Right) Radar chart of leaderboard results.

Image Examples

MY ALT TEXT

Benchmark Comparison

MY ALT TEXT

Comparisons between AlignMMBench and other benchmarks. Multi Category represent whether the categories of questions in benchmark are greater than 2. Dialogue Context represent whether the benchmark contains multi-turn dialogue scenarios.

Evaluating CritiqueVLM

MY ALT TEXT

Mean absolute error (e), Pearson (r), Spearman (ρ), Kendall (τ) correlations, Fuzzy division scores (f) and Strict division scores (s) are calculated on the internal and external test sets. The best performance among the methods is bold. These results demonstrates that CritiqueVLM achieves superior performance with only 6 billion parameters.

BibTeX

@misc{wu2024alignmmbench,
        title={AlignMMBench: Evaluating Chinese Multimodal Alignment in Large Vision-Language Models}, 
        author={Yuhang Wu and Wenmeng Yu and Yean Cheng and Yan Wang and Xiaohan Zhang and Jiazheng Xu and Ming Ding and Yuxiao Dong},
        year={2024},
        eprint={2406.09295},
        archivePrefix={arXiv}
  }