Saved in:
| Main Authors: | , , , , , , , , , , , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2502.10391 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866912232222228480 |
|---|---|
| author | Zhang, Yi-Fan Yu, Tao Tian, Haochen Fu, Chaoyou Li, Peiyan Zeng, Jianshu Xie, Wulin Shi, Yang Zhang, Huanyu Wu, Junkang Wang, Xue Hu, Yibo Wen, Bin Yang, Fan Zhang, Zhang Gao, Tingting Zhang, Di Wang, Liang Jin, Rong Tan, Tieniu |
| author_facet | Zhang, Yi-Fan Yu, Tao Tian, Haochen Fu, Chaoyou Li, Peiyan Zeng, Jianshu Xie, Wulin Shi, Yang Zhang, Huanyu Wu, Junkang Wang, Xue Hu, Yibo Wen, Bin Yang, Fan Zhang, Zhang Gao, Tingting Zhang, Di Wang, Liang Jin, Rong Tan, Tieniu |
| contents | Despite notable advancements in Multimodal Large Language Models (MLLMs), most state-of-the-art models have not undergone thorough alignment with human preferences. This gap exists because current alignment research has primarily achieved progress in specific areas (e.g., hallucination reduction), while the broader question of whether aligning models with human preferences can systematically enhance MLLM capability remains largely unexplored. To this end, we introduce MM-RLHF, a dataset containing $\mathbf{120k}$ fine-grained, human-annotated preference comparison pairs. This dataset represents a substantial advancement over existing resources, offering superior size, diversity, annotation granularity, and quality. Leveraging this dataset, we propose several key innovations to improve both the quality of reward models and the efficiency of alignment algorithms. Notably, we introduce a Critique-Based Reward Model, which generates critiques of model outputs before assigning scores, offering enhanced interpretability and more informative feedback compared to traditional scalar reward mechanisms. Additionally, we propose Dynamic Reward Scaling, a method that adjusts the loss weight of each sample according to the reward signal, thereby optimizing the use of high-quality comparison pairs. Our approach is rigorously evaluated across $\mathbf{10}$ distinct dimensions and $\mathbf{27}$ benchmarks, with results demonstrating significant and consistent improvements in model performance. Specifically, fine-tuning LLaVA-ov-7B with MM-RLHF and our alignment algorithm leads to a $\mathbf{19.5}$% increase in conversational abilities and a $\mathbf{60}$% improvement in safety.
We have open-sourced the preference dataset, reward model, training and evaluation code, as well as reward modeling and safety benchmarks. For more details, please visit our project page: https://mm-rlhf.github.io. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2502_10391 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | MM-RLHF: The Next Step Forward in Multimodal LLM Alignment Zhang, Yi-Fan Yu, Tao Tian, Haochen Fu, Chaoyou Li, Peiyan Zeng, Jianshu Xie, Wulin Shi, Yang Zhang, Huanyu Wu, Junkang Wang, Xue Hu, Yibo Wen, Bin Yang, Fan Zhang, Zhang Gao, Tingting Zhang, Di Wang, Liang Jin, Rong Tan, Tieniu Computation and Language Computer Vision and Pattern Recognition Despite notable advancements in Multimodal Large Language Models (MLLMs), most state-of-the-art models have not undergone thorough alignment with human preferences. This gap exists because current alignment research has primarily achieved progress in specific areas (e.g., hallucination reduction), while the broader question of whether aligning models with human preferences can systematically enhance MLLM capability remains largely unexplored. To this end, we introduce MM-RLHF, a dataset containing $\mathbf{120k}$ fine-grained, human-annotated preference comparison pairs. This dataset represents a substantial advancement over existing resources, offering superior size, diversity, annotation granularity, and quality. Leveraging this dataset, we propose several key innovations to improve both the quality of reward models and the efficiency of alignment algorithms. Notably, we introduce a Critique-Based Reward Model, which generates critiques of model outputs before assigning scores, offering enhanced interpretability and more informative feedback compared to traditional scalar reward mechanisms. Additionally, we propose Dynamic Reward Scaling, a method that adjusts the loss weight of each sample according to the reward signal, thereby optimizing the use of high-quality comparison pairs. Our approach is rigorously evaluated across $\mathbf{10}$ distinct dimensions and $\mathbf{27}$ benchmarks, with results demonstrating significant and consistent improvements in model performance. Specifically, fine-tuning LLaVA-ov-7B with MM-RLHF and our alignment algorithm leads to a $\mathbf{19.5}$% increase in conversational abilities and a $\mathbf{60}$% improvement in safety. We have open-sourced the preference dataset, reward model, training and evaluation code, as well as reward modeling and safety benchmarks. For more details, please visit our project page: https://mm-rlhf.github.io. |
| title | MM-RLHF: The Next Step Forward in Multimodal LLM Alignment |
| topic | Computation and Language Computer Vision and Pattern Recognition |
| url | https://arxiv.org/abs/2502.10391 |