A new paper on arXiv introduces MHR-Router, a system for routing image-question queries to the most suitable vision-language model (VLM). The authors note that VLMs have complementary capabilities, but choosing the right expert for a given query is difficult because inputs are heterogeneous and model strengths vary.
MHR-Router stands for Multimodal Heterogeneous Retrieval-Enhanced Adaptive Routing. The approach combines retrieval mechanisms with adaptive routing, likely using retrieved multimodal information to better match a query to a model. The abstract emphasizes the challenge of heterogeneous multimodal inputs and varying expert strengths, which the router aims to address.
The work is at an early stage, with the abstract providing a high-level description. No experimental results are included in the abstract, so the effectiveness of the proposed routing method is not yet detailed. The paper is available on arXiv under identifier 2610.13278.