Visual Instruction Bottleneck Tuning

TL;DR

This paper introduces Visual Instruction Bottleneck Tuning (Vittle), a novel method based on the information bottleneck principle, to improve the robustness of multimodal large language models under distribution shifts without requiring more data or larger models.

Contribution

Vittle applies the information bottleneck principle to enhance MLLM generalization, providing a theoretical foundation and practical implementation that improves robustness across diverse tasks and datasets.

Findings

Vittle improves MLLM robustness under distribution shifts.

Vittle outperforms baseline methods on 45 datasets including 30 shift scenarios.

Theoretical link between Vittle and information-theoretic robustness metrics.

Abstract

Despite widespread adoption, multimodal large language models (MLLMs) suffer performance degradation when encountering unfamiliar queries under distribution shifts. Existing methods to improve MLLM generalization typically require either more instruction data or larger advanced model architectures, both of which incur non-trivial human labor or computational costs. In this work, we take an alternative approach to enhance the generalization and robustness of MLLMs under distribution shifts, from a representation learning perspective. Inspired by information bottleneck (IB) principle, we derive a variational lower bound of the IB for MLLMs and devise a practical implementation, Visual Instruction Bottleneck Tuning (Vittle). We then provide a theoretical justification of Vittle by revealing its connection to an information-theoretic robustness metric of MLLM. Empirical validation of multiple MLLMs on open-ended and closed-form question answering and object hallucination detection tasks over 45 datasets, including 30 shift scenarios, demonstrates that Vittle consistently improves the MLLM's robustness under shifts by pursuing the learning of a minimal sufficient representation.