X-LLM: Bootstrapping Advanced Large Language Models by Treating Multi-Modalities as Foreign Languages

TL;DR

X-LLM introduces a novel approach to enable large language models to handle multiple modalities by converting them into language-like representations, enhancing multimodal capabilities without modifying the core LLM architecture.

Contribution

The paper proposes a new framework, X-LLM, that aligns multimodal inputs with a frozen LLM using X2L interfaces, allowing multimodal integration through a three-stage training process.

Findings

X-LLM achieves multimodal chat abilities comparable to GPT-4 on unseen data.

It attains 84.5% relative score on a synthetic multimodal instruction dataset.

Demonstrates potential for LLM-based speech recognition and multimodal understanding.

Abstract

Large language models (LLMs) have demonstrated remarkable language abilities. GPT-4, based on advanced LLMs, exhibits extraordinary multimodal capabilities beyond previous visual language models. We attribute this to the use of more advanced LLMs compared with previous multimodal models. Unfortunately, the model architecture and training strategies of GPT-4 are unknown. To endow LLMs with multimodal capabilities, we propose X-LLM, which converts Multi-modalities (images, speech, videos) into foreign languages using X2L interfaces and inputs them into a large Language model (ChatGLM). Specifically, X-LLM aligns multiple frozen single-modal encoders and a frozen LLM using X2L interfaces, where ``X'' denotes multi-modalities such as image, speech, and videos, and ``L'' denotes languages. X-LLM's training consists of three stages: (1) Converting Multimodal Information: The first stage trains each X2L interface to align with its respective single-modal encoder separately to convert multimodal information into languages. (2) Aligning X2L representations with the LLM: single-modal encoders are aligned with the LLM through X2L interfaces independently. (3) Integrating multiple modalities: all single-modal encoders are aligned with the LLM through X2L interfaces to integrate multimodal capabilities into the LLM. Our experiments show that X-LLM demonstrates impressive multimodel chat abilities, sometimes exhibiting the behaviors of multimodal GPT-4 on unseen images/instructions, and yields a 84.5\% relative score compared with GPT-4 on a synthetic multimodal instruction-following dataset. And we also conduct quantitative tests on using LLM for ASR and multimodal ASR, hoping to promote the era of LLM-based speech recognition.