[Paper Review] Reformulating Vision-Language Foundation Models and Datasets Towards Universal Multimodal Assistants

Tianyu Yu, Jinyi Hu+11 more|arXiv (Cornell University)|Oct 1, 2023

Multimodal Machine Learning Applications4 citations

TL;DR

This paper proposes Muffin, a novel framework that uses pre-trained vision-language models as 'out-of-the-box' bridges between vision encoders and large language models, eliminating the need for additional feature-alignment training. It also introduces UniMM-Chat, a 1.1M-sample multimodal instruction dataset created by merging complementary annotations from diverse datasets, which enables state-of-the-art performance across vision-language tasks, outperforming models like LLaVA and InstructBLIP.

ABSTRACT

Recent Multimodal Large Language Models (MLLMs) exhibit impressive abilities to perceive images and follow open-ended instructions. The capabilities of MLLMs depend on two crucial factors: the model architecture to facilitate the feature alignment of visual modules and large language models; the multimodal instruction tuning datasets for human instruction following. (i) For the model architecture, most existing models introduce an external bridge module to connect vision encoders with language models, which needs an additional feature-alignment pre-training. In this work, we discover that compact pre-trained vision language models can inherently serve as ``out-of-the-box'' bridges between vision and language. Based on this, we propose Muffin framework, which directly employs pre-trained vision-language models to act as providers of visual signals. (ii) For the multimodal instruction tuning datasets, existing methods omit the complementary relationship between different datasets and simply mix datasets from different tasks. Instead, we propose UniMM-Chat dataset which explores the complementarities of datasets to generate 1.1M high-quality and diverse multimodal instructions. We merge information describing the same image from diverse datasets and transforms it into more knowledge-intensive conversation data. Experimental results demonstrate the effectiveness of the Muffin framework and UniMM-Chat dataset. Muffin achieves state-of-the-art performance on a wide range of vision-language tasks, significantly surpassing state-of-the-art models like LLaVA and InstructBLIP. Our model and dataset are all accessible at https://github.com/thunlp/muffin.

Motivation & Objective

• To address the inefficiency and complexity of existing MLLM architectures that require additional alignment modules or re-samplers.
• To overcome the limitations of current multimodal instruction tuning datasets, which often lack depth, diversity, and complementary information.
• To develop a unified, high-quality, and knowledge-intensive multimodal instruction dataset that enhances reasoning and factual grounding in LLMs.
• To demonstrate that direct integration of pre-trained VLMs with LLMs can achieve superior performance without extra alignment training.
• To establish a new paradigm for constructing multimodal datasets by fusing annotations from multiple sources for the same image.

Proposed method

• Muffin directly uses compact pre-trained vision-language models (e.g., ALBEF, CoCa, BEiT-3) as visual signal providers, bypassing the need for external bridge modules or re-samplers.
• It introduces learnable query vectors in the VLM’s embedding space to extract and inject visual representations into the LLM, enabling end-to-end training.
• The UniMM-Chat dataset is constructed by merging annotations from multiple vision-language datasets (e.g., COCO) for the same image, creating richer, more knowledge-intensive conversation data.
• A flexible prompting strategy using GPT-4 is applied to transform merged annotations into diverse, high-quality, multi-turn dialogues with detailed responses.
• The framework supports training with high-resolution images during pre-training and instruction tuning to preserve fine-grained visual details.
• A new benchmark, UniMM-Bench, is introduced to evaluate multimodal reasoning and world knowledge, using GPT-4 to score model outputs on complex, open-ended questions.

Experimental results

Research questions

• RQ1 Can pre-trained vision-language models serve as effective, out-of-the-box bridges between vision encoders and large language models without additional alignment modules?
• RQ2 How does merging complementary annotations from multiple datasets improve the quality and knowledge density of multimodal instruction tuning data?
• RQ3 To what extent does training on a knowledge-intensive, multi-source annotated dataset like UniMM-Chat improve reasoning and factual accuracy in MLLMs?
• RQ4 Can a simpler, unified training framework like Muffin outperform complex architectures that rely on external re-samplers or projectors?
• RQ5 What impact does image resolution and fine-tuning strategy have on the performance of multimodal models in complex reasoning tasks?

Key findings

• Muffin achieves state-of-the-art performance across multiple vision-language benchmarks, outperforming LLaVA and InstructBLIP on Visual Question Answering, Visual Chat, and detail description tasks.
• On the UniMM-Bench benchmark, Muffin achieves a 73.4% score, significantly outperforming models fine-tuned on original datasets or less comprehensive variants.
• The model trained on UniMM-Chat achieves 85.7% accuracy on the LLaVA test set, demonstrating strong generalization and robustness to out-of-domain examples.
• Ablation studies show that using merged annotations (UniMM-Chat) leads to better performance than using isolated or non-complementary data (e.g., UniMM-Chat-sep), confirming the value of data fusion.
• Reducing image resolution during pre-training or fine-tuning leads to a significant drop in performance, highlighting the importance of high-resolution visual input for complex tasks.
• Freezing the LLM during instruction tuning results in a 10-point drop in performance on UniMM-Bench, indicating that fine-tuning the LLM is essential for optimal knowledge integration.