LLaVA Steering: Visual Instruction Tuning with 500x Fewer Parameters through Modality Linear Representation-Steering

TL;DR

This paper introduces MoReS, a linear transformation method that re-balances visual and textual modalities in multimodal models, drastically reducing trainable parameters while maintaining performance in visual tasks.

Contribution

The paper proposes MoReS, a novel linear representation-steering technique that significantly decreases the number of trainable parameters needed for visual instruction tuning.

Findings

MoReS reduces trainable parameters by 500x compared to LoRA.

LLaVA Steering models achieve comparable performance with fewer parameters.

The platform enables quick customization and evaluation of multimodal models.

Abstract

Multimodal Large Language Models (MLLMs) have significantly advanced visual tasks by integrating visual representations into large language models (LLMs). The textual modality, inherited from LLMs, equips MLLMs with abilities like instruction following and in-context learning. In contrast, the visual modality enhances performance in downstream tasks by leveraging rich semantic content, spatial information, and grounding capabilities. These intrinsic modalities work synergistically across various visual tasks. Our research initially reveals a persistent imbalance between these modalities, with text often dominating output generation during visual instruction tuning. This imbalance occurs when using both full fine-tuning and parameter-efficient fine-tuning (PEFT) methods. We then found that re-balancing these modalities can significantly reduce the number of trainable parameters required, inspiring a direction for further optimizing visual instruction tuning. We introduce Modality Linear Representation-Steering (MoReS) to achieve the goal. MoReS effectively re-balances the intrinsic modalities throughout the model, where the key idea is to steer visual representations through linear transformations in the visual subspace across each model layer. To validate our solution, we composed LLaVA Steering, a suite of models integrated with the proposed MoReS method. Evaluation results show that the composed LLaVA Steering models require, on average, 500 times fewer trainable parameters than LoRA needs while still achieving comparable performance across three visual benchmarks and eight visual question-answering tasks. Last, we present the LLaVA Steering Factory, an in-house developed platform that enables researchers to quickly customize various MLLMs with component-based architecture for seamlessly integrating state-of-the-art models, and evaluate their intrinsic modality imbalance.