Multi-modal Preference Alignment Remedies Degradation of Visual Instruction Tuning on Language Models

Anonymous

Abstract: In production, multi-modal large language models (MLLMs) are expected to support multi-turn queries of interchanging image and text modalitie. However, the current MLLMs trained with visual-question-answering (VQA) datasets could suffer from degradation, as VQA datasets lack the diversity and complexity of the original text instruction datasets which the underlying language model had been trained with. To address this challenging degradation, we first collect a lightweight (6k entries) VQA preference dataset where answers were annotated by Gemini for 5 quality metrics in a granular fashion, and investigate standard Supervised Fine-tuning, rejection sampling, Direct Preference Optimization (DPO), and SteerLM. Our findings indicate that the with DPO we are able to surpass instruction-following capabilities of the language model, achieving a 6.73 score on MT-Bench, compared to Vicuna's 6.57 and LLaVA's 5.99 despite small data scale. This enhancement in textual instruction proficiency correlates with boosted visual instruction performance (+4.9\% on MM-Vet, +6\% on LLaVA-Bench), with minimal alignment tax on visual knowledge benchmarks compared to previous RLHF approach. In conclusion, we propose a distillation-based multi-modal alignment model with fine-grained annotations on a small dataset that reconciles the textual and visual performance of MLLMs, restoring and boosting language capability after visual instruction tuning.

Paper Type: long

Research Area: Multimodality and Language Grounding to Vision, Robotics and Beyond

Contribution Types: NLP engineering experiment, Data resources

Languages Studied: English