LLaVA-RE: Binary Image-Text Relevancy Evaluation with Multimodal Large Language Model

Multimodal generative AI usually involves generating image or text responses given inputs in another modality. The evaluation of image-text relevancy is essential for measuring response quality or ranking candidate responses. In particular, binary relevancy evaluation, i.e., ``Relevant'' vs. ``Not Relevant'', is a fundamental problem. However, this is a challenging task considering that texts have diverse formats and the definition of relevancy varies in different scenarios. We find that Multimodal Large Language Models (MLLMs) are an ideal choice to build such evaluators, as they can flexibly handle complex text formats and take in additional task information. In this paper, we present LLaVA-RE, a first attempt for binary image-text relevancy evaluation with MLLM. It follows the LLaVA architecture and adopts detailed task instructions and multimodal in-context samples. In addition, we propose a novel binary relevancy data set that covers various tasks. Experimental results validate the effectiveness of our framework.

RS-DPO: A Hybrid Rejection Sampling and Direct Preference Optimization Method for Alignment of Large Language Models

Reinforcement learning from human feedback (RLHF) has been extensively employed to align large language models with user intent. However, proximal policy optimization (PPO) based RLHF is occasionally unstable requiring significant hyperparameter finetuning, and computationally expensive to maximize the estimated reward during alignment. Recently, direct preference optimization (DPO) is proposed to address those challenges. However, DPO relies on contrastive responses generated from human annotator and alternative LLM, instead of the policy model, limiting the effectiveness of the RLHF. In this paper, we addresses both challenges by systematically combining rejection sampling (RS) and DPO. Our proposed method, RS-DPO, initiates with the development of a supervised fine-tuned policy model (SFT). A varied set of k responses per prompt are sampled directly from the SFT model. RS-DPO identifies pairs of contrastive samples based on their reward distribution. Finally, we apply DPO with the contrastive samples to align the model to human preference. Our experiments indicate that our proposed method effectively fine-tunes LLMs with limited resource environments, leading to improved alignment with user intent. Furthermore, it outperforms existing methods, including RS, PPO, and DPO.