TinyAlign: Boosting Lightweight Vision-Language Models by Mitigating Modal Alignment Bottlenecks

2025-05-19

AI Summary

Lightweight vision-language models (VLMs) suffer from a cross-modal alignment bottleneck due to the limited representational capacity of their language encoders. Method: This paper is the first to attribute this bottleneck to insufficient effective mutual information (EMI) between modalities. We propose TinyAlign—a retrieval-augmented generation (RAG)-based framework that constructs an updatable multimodal memory bank and employs a lightweight connector to enable dynamic contextual injection and alignment optimization. The method integrates EMI-theoretic analysis, efficient memory retrieval, and parameter-efficient fine-tuning. Contribution/Results: TinyAlign substantially reduces training loss and accelerates convergence. It achieves baseline performance using only 40% of the fine-tuning data, consistently improving alignment quality and generalization across multiple downstream tasks. The approach significantly enhances data efficiency and adaptability to resource-constrained settings, without increasing model size or inference latency.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Large Vision Models
• Natural Language Processing: Language Grounding & Multi-modal NLP

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Lightweight Vision-Language Models (VLMs) are indispensable for resource-constrained applications. The prevailing approach to aligning vision and language models involves freezing both the vision encoder and the language model while training small connector modules. However, this strategy heavily depends on the intrinsic capabilities of the language model, which can be suboptimal for lightweight models with limited representational capacity. In this work, we investigate this alignment bottleneck through the lens of mutual information, demonstrating that the constrained capacity of the language model inherently limits the Effective Mutual Information (EMI) between multimodal inputs and outputs, thereby compromising alignment quality. To address this challenge, we propose TinyAlign, a novel framework inspired by Retrieval-Augmented Generation, which strategically retrieves relevant context from a memory bank to enrich multimodal inputs and enhance their alignment. Extensive empirical evaluations reveal that TinyAlign significantly reduces training loss, accelerates convergence, and enhances task performance. Remarkably, it allows models to achieve baseline-level performance with only 40% of the fine-tuning data, highlighting exceptional data efficiency. Our work thus offers a practical pathway for developing more capable lightweight VLMs while introducing a fresh theoretical lens to better understand and address alignment bottlenecks in constrained multimodal systems.

Problem

• Lightweight VLMs face modal alignment bottlenecks due to limited representational capacity.
• Constrained language models limit Effective Mutual Information in multimodal alignment.
• TinyAlign enhances alignment by retrieving context to enrich multimodal inputs.

Innovation

• Uses mutual information to analyze alignment bottlenecks
• Introduces TinyAlign with retrieval-augmented context enrichment
• Achieves high data efficiency with 40% fine-tuning data