Assess PartnerMAS performance with lightweight or open-source LLM backbones

Determine the performance of the PartnerMAS hierarchical multi-agent framework for business partner selection on high-dimensional, heterogeneous tabular features when implemented with lighter or open-source large language model backbones instead of advanced GPT backbones, particularly in resource-constrained environments.

Background

PartnerMAS is a hierarchical multi-agent system designed for business partner selection in high-dimensional settings, decomposing evaluation into Planner, Specialized, and Supervisor agents. The study benchmarks performance on a curated venture capital co-investment dataset and reports that PartnerMAS consistently outperforms single-agent and debate-style baselines in match rate and efficiency.

However, the evaluation is primarily conducted with advanced GPT backbones. The authors explicitly state that it remains open how PartnerMAS performs when using lighter or open-source models, which may be necessary for deployment in resource-constrained environments. Understanding this performance is important for broader applicability, cost-effectiveness, and potential use with open models.