M3PO: Multimodal-Model-Guided Preference Optimization for Visual Instruction Following

Large Vision-Language Models (LVLMs) hold immense potential for complex multimodal instruction following, yet their development is often hindered by the high cost and inconsistency of human annotation required for effective fine-tuning and preference alignment. Traditional supervised fine-tuning (SFT) and existing preference optimization methods like RLHF and DPO frequently struggle to efficiently leverage the model's own generation space to identify highly informative "hard negative" samples. To address these challenges, we propose Multimodal-Model-Guided Preference Optimization (M3PO), a novel and data-efficient method designed to enhance LVLMs' capabilities in visual instruction following. M3PO intelligently selects the most "learning-valuable" preference sample pairs from a diverse pool of LVLM-generated candidates. This selection is driven by a sophisticated mechanism that integrates two crucial signals: a Multimodal Alignment Score (MAS) to assess external quality and the model's Self-Consistency / Confidence (log-probability) to gauge internal belief. These are combined into a novel M3P-Score, which specifically identifies preferred responses and challenging dispreferred responses that the model might confidently generate despite being incorrect. These high-quality preference pairs are then used for efficient Direct Preference Optimization (DPO) fine-tuning on base LVLMs like LLaVA-1.5 (7B/13B) using LoRA. Our extensive experiments demonstrate that M3PO consistently outperforms strong baselines, including SFT, simulated RLHF, vanilla DPO, and RM-DPO, across a comprehensive suite of multimodal instruction following benchmarks (MME-Bench, POPE, IFT, Human Pref. Score).