Instruction-Oriented Preference Alignment for Enhancing Multi-Modal Comprehension Capability of MLLMs
This work addresses the enhancement of multi-modal comprehension in MLLMs, which is important for applications requiring accurate interpretation of visual and textual data, though it appears incremental as it builds on existing preference alignment strategies.
The paper tackles the problem that existing preference alignment methods for Multimodal Large Language Models (MLLMs) focus on hallucination mitigation but neglect multi-modal comprehension capabilities, proposing Instruction-oriented Preference Alignment (IPA) to automatically construct alignment preferences based on instruction fulfillment, which improves performance on benchmarks like hallucination evaluation, visual question answering, and text understanding tasks using Qwen2VL-7B.
Preference alignment has emerged as an effective strategy to enhance the performance of Multimodal Large Language Models (MLLMs) following supervised fine-tuning. While existing preference alignment methods predominantly target hallucination factors, they overlook the factors essential for multi-modal comprehension capabilities, often narrowing their improvements on hallucination mitigation. To bridge this gap, we propose Instruction-oriented Preference Alignment (IPA), a scalable framework designed to automatically construct alignment preferences grounded in instruction fulfillment efficacy. Our method involves an automated preference construction coupled with a dedicated verification process that identifies instruction-oriented factors, avoiding significant variability in response representations. Additionally, IPA incorporates a progressive preference collection pipeline, further recalling challenging samples through model self-evolution and reference-guided refinement. Experiments conducted on Qwen2VL-7B demonstrate IPA's effectiveness across multiple benchmarks, including hallucination evaluation, visual question answering, and text understanding tasks, highlighting its capability to enhance general comprehension.