Researchers unveil HugSelect for transparent AI model selection

A team of researchers from Utrecht University, Shiraz University, and Wageningen University & Research has developed HugSelect, a novel framework designed to tackle the growing challenge of selecting foundation models (FMs) as reusable software components. Unlike traditional model hubs that rely on popularity metrics or opaque recommendations, HugSelect treats model selection as an explicit, auditable software-engineering task.

The framework aggregates data from 71,274 models by combining repository metadata, extracted functional capabilities (e.g., multilingual support, task-specific features), and perceived quality attributes derived from community discussions. It then ranks candidates using a weighted additive model that exposes criterion-level score decompositions, enabling users to understand *why* a model was recommended. In evaluations against four commercial LLM-based recommendation systems across 44 scenarios, HugSelect achieved a model-level Coverage@10 of 0.61 and a family-level Coverage@10 of 0.91—comparable to commercial solutions—while providing stable, traceable, and inspectable reasoning. The extraction pipelines demonstrated strong performance, with an F1 score of 0.801 for functional features and 0.84 accuracy for quality-attribute mapping.