Self-Captioning Multimodal Interaction Tuning: Amplifying Exploitable Redundancies for Robust Vision Language Models

2026-05-03

AI Summary

This work addresses the susceptibility of vision-language models to hallucinations and insufficient robustness under modality ambiguity or corruption. It presents the first systematic approach to enhancing model reliability by leveraging multimodal redundancy. The proposed method introduces a multimodal interaction gating mechanism that, within a self-descriptive workflow, transforms modality-unique information into redundant representations. Coupled with techniques for multimodal information decomposition and reconstruction, this framework strengthens the model’s reliance on consistent cross-modal signals. Experimental results demonstrate substantial improvements in robustness: visual-induced errors are reduced by 38.3% and model consistency increases by 16.8% on standard benchmarks.

Abstract

Current vision language models face hallucination and robustness issues against ambiguous or corrupted modalities. We hypothesize that these issues can be addressed by exploiting the shared information between modalities to compensate for the impaired one. To this end, we analyze multimodal interactions -- redundant (shared), unique (exclusive), and synergistic (emergent) task-relevant information provided by the modalities -- to determine their impacts on model reliability. Specifically, amplifying redundant interactions would increase this exploitable shared information to resolve these issues; yet, modern instruction datasets often eliminate redundancies to prioritize visual grounding. We bridge this gap through a self-captioning workflow featuring a \textsc{Multimodal Interaction Gate}: a mechanism to convert unique interactions into redundant interactions. Our findings suggest that increasing redundancy can reduce visual induced errors by 38.3\% and improve consistency by 16.8\%.

Problem

• vision language models
• hallucination
• robustness
• multimodal interaction
• redundancy

Innovation

• multimodal interaction
• redundancy amplification
• vision-language models
• self-captioning
• robustness