New HIVE framework improves vision-language AI by using hierarchical cross-attention to better integrate visual features with large language models

HIVE (Hierarchical Pre-Training of Vision Encoders) introduces hierarchical cross-attention between vision encoders and LLMs instead of treating them as independent modules

The framework maintains structured feature fusion across multiple layers rather than flattening image embeddings, enabling better gradient flow and representation learning

A three-stage progressive training strategy aligns the vision encoder with the LLM for stable optimization and effective multimodal fusion

Empirical evaluations show the approach enhances vision-language alignment compared to conventional methods