Optimal Feature Embedding for Document Large Visual Language Model

摘要

Document Large Vision Language Models excel in document-centric tasks and have become a key focus of research. Existing frameworks embed features from a lightweight, document-specific encoder into the first layer of a general-purpose Vision Language Model (VLM). However, this introduces a feature mismatch problem. VLMs typically consist of many stacked layers, with the feature hierarchy becoming increasingly abstract at higher layers. Specifically, the first-layer feature in a VLM is token-level, whereas the feature from the encoder is task-level, resulting in a mismatch. Consequently, it is crucial to identify an optimal layer within the VLM for embedding the encoder's features. Inspired by physics, we reformulate the search for the optimal embedding as a problem of finding the shortest time curve. Leveraging the properties of the shortest time curve, we theoretically derive a task-agnostic proxy score that requires only partial training and propose our searching framework, Brac4VLM. Our theoretical derivation shows that Brac4VLM reduces search time by 97.8% compared to brute-force methods. Experimental results further demonstrate that Brac4VLM identifies embedding points that closely align with the true optima. Moreover, the DocVLM with the optimal embedding position identified achieves state-of-the-art performance across various document-centric tasks. Codes: https://github.com/MaxKinny/Brac4VLM.