Multimodal RewardBench: Holistic Evaluation of Reward Models for Vision Language Models
This provides a challenging testbed for researchers developing reward models in vision-language models, though it is incremental as it focuses on benchmarking rather than novel methods.
The authors tackled the lack of comprehensive benchmarks for evaluating multimodal reward models in vision-language models by introducing Multimodal RewardBench, an expert-annotated dataset covering six domains, and found that top-performing models like Gemini 1.5 Pro and Claude 3.5 Sonnet achieve only 72% overall accuracy, with particular struggles in reasoning and safety.
Reward models play an essential role in training vision-language models (VLMs) by assessing output quality to enable aligning with human preferences. Despite their importance, the research community lacks comprehensive open benchmarks for evaluating multimodal reward models in VLMs. To address this gap, we introduce Multimodal RewardBench, an expert-annotated benchmark covering six domains: general correctness, preference, knowledge, reasoning, safety, and visual question-answering. Our dataset comprises 5,211 annotated (prompt, chosen response, rejected response) triplets collected from various VLMs. In evaluating a range of VLM judges, we find that even the top-performing models, Gemini 1.5 Pro and Claude 3.5 Sonnet, achieve only 72% overall accuracy. Notably, most models struggle in the reasoning and safety domains. These findings suggest that Multimodal RewardBench offers a challenging testbed for advancing reward model development across multiple domains. We release the benchmark at https://github.com/facebookresearch/multimodal_rewardbench.