Title: CUPID in the Model Zoo: Online Matchmaking for Selecting Your Dream LLM

Original: arXiv:2606.00846v1 Announce Type: new

Abstract: As the landscape of Large Language Models (LLMs) expands rapidly, users are increasingly tasked with the difficult job of identifying the most suitable model for specific applications. This selection process is complicated by the fact that each model possesses unique, yet often hidden, latent characteristics. Furthermore, many users struggle to define or communicate the specific traits they prioritize in model outputs or deployment scenarios. To address these hurdles, we present an active learning framework designed for interaction efficiency. This system employs a dueling bandit algorithm to continuously pick pairs of LLMs, gather user feedback on their respective responses, and refine its understanding of the user’s underlying preferences. We also introduce a new belief-aware upper confidence bound approach that effectively navigates the trade-off between exploring the broader model catalog and exploiting the preferences we have inferred. This strategy facilitates a swift and accurate alignment between user requirements and model capabilities, even when constrained by specific time and cost limits. Our experiments, which include tests on various LLMs and human subject studies, demonstrate that our approach successfully pairs users with well-suited models more economically and efficiently than current methods.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC