Hierarchical Pre-Training of Vision Encoders with Large Language Models

TL;DR

HIVE introduces hierarchical cross-attention for vision encoders and LLMs, enabling structured feature fusion that improves multimodal learning and performance across vision-language tasks.

Contribution

The paper presents a novel hierarchical pre-training framework that enhances vision-language alignment through structured feature fusion and a three-stage training strategy.

Findings

HIVE outperforms self-attention-based methods on multiple benchmarks.

Hierarchical feature integration improves gradient flow and representation learning.

Effective multimodal fusion enhances performance in image classification and vision-language tasks.

Abstract

The field of computer vision has experienced significant advancements through scalable vision encoders and multimodal pre-training frameworks. However, existing approaches often treat vision encoders and large language models (LLMs) as independent modules, limiting the integration of hierarchical visual features. In this work, we propose HIVE (Hierarchical Pre-Training of Vision Encoders), a novel framework that enhances vision-language alignment by introducing hierarchical cross-attention between the vision encoder and LLM. Unlike conventional methods that flatten image embeddings, HIVE enables structured feature fusion across multiple layers, improving gradient flow and representation learning. To optimize this interaction, we introduce a three-stage training strategy that progressively aligns the vision encoder with the LLM, ensuring stable optimization and effective multimodal fusion. Empirical evaluations demonstrate that HIVE achieves superior performance not only in image classification but also on various vision-language tasks, outperforming self-attention-based methods in benchmarks such as MME, GQA, OK-VQA, and ScienceQA. Our results highlight the benefits of hierarchical feature integration, paving the way for more efficient and expressive vision-language models.