Strengthening Multimodal Large Language Model with Bootstrapped Preference Optimization

TL;DR

This paper introduces Bootstrapped Preference Optimization (BPO), a novel method that reduces pretrained bias in Multimodal Large Language Models, leading to improved grounding in visual inputs and state-of-the-art performance.

Contribution

The paper proposes a new preference learning approach using bootstrapped negative responses to mitigate pretrained bias in MLLMs, enhancing their visual grounding capabilities.

Findings

Significant performance improvements across multiple benchmarks.

Effective suppression of pretrained bias in MLLMs.

Enhanced grounding in visual inputs.

Abstract

Multimodal Large Language Models (MLLMs) excel in generating responses based on visual inputs. However, they often suffer from a bias towards generating responses similar to their pretraining corpus, overshadowing the importance of visual information. We treat this bias as a "preference" for pretraining statistics, which hinders the model's grounding in visual input. To mitigate this issue, we propose Bootstrapped Preference Optimization (BPO), which conducts preference learning with datasets containing negative responses bootstrapped from the model itself. Specifically, we propose the following two strategies: 1) using distorted image inputs to the MLLM for eliciting responses that contain signified pretraining bias; 2) leveraging text-based LLM to explicitly inject erroneous but common elements into the original response. Those undesirable responses are paired with original annotated responses from the datasets to construct the preference dataset, which is subsequently utilized to perform preference learning. Our approach effectively suppresses pretrained LLM bias, enabling enhanced grounding in visual inputs. Extensive experimentation demonstrates significant performance improvements across multiple benchmarks, advancing the state-of-the-art in multimodal conversational systems.