Continual Model Routing in Evolving Model Hubs

Researchers have proposed two new models to improve efficiency in AI model routing and table reasoning, addressing challenges posed by the rapid growth of pre-trained models in AI model hubs.

The first model, CARvE, is a contrastive embedding approach for efficient continual model routing in evolving model hubs. It uses checkpoint-based anchoring and structured replay to outperform zero-shot retrieval, fine-tuning, and adapter-merging baselines in model, family, and domain-level accuracy^[1]. AI model hubs provide access to a rapidly growing collection of pre-trained models, posing challenges in scaling model selection and updating routing mechanisms. To address this, researchers formalised the setting as Continual Model Routing (CMR) and proposed CMRBench, a large-scale benchmark simulating realistic hub expansion with over 2,000 candidate models^[1]. Meanwhile, a separate research effort introduced EcoTab, a table-aware stepwise routing framework for efficient table reasoning. EcoTab estimates the uncertainties of table tokens and text tokens separately and combines them for routing, achieving a better balance between accuracy and efficiency. This is necessary as Large Reasoning Models (LRMs) incur substantial inference cost due to long reasoning traces, and existing methods fail to model separately the uncertainty of table tokens and text tokens^[2].