PartnerMAS: An LLM Hierarchical Multi-Agent Framework for Business Partner Selection on High-Dimensional Features

TL;DR

PartnerMAS addresses the challenge of high-dimensional business partner selection by decomposing the task into a Planner, Specialist, and Supervisor within a hierarchical multi-agent system. The approach is validated on a VC co-investment benchmark of 140 cases, showing consistent gains over single-agent and debate-based baselines, with up to 10–15 percentage points higher match rates and improved cost-efficiency. The work provides a new benchmark, demonstrates how role specialization and structured aggregation outperform mere model scaling, and offers detailed analyses of how planning, specialization, and supervision contribute to robust decision-making in data-rich domains. The framework holds promise for broad application to similar high-dimensional decision tasks in finance and beyond, while acknowledging data-, model-, and supervision-related limitations that warrant further research.

Abstract

High-dimensional decision-making tasks, such as business partner selection, involve evaluating large candidate pools with heterogeneous numerical, categorical, and textual features. While large language models (LLMs) offer strong in-context reasoning capabilities, single-agent or debate-style systems often struggle with scalability and consistency in such settings. We propose PartnerMAS, a hierarchical multi-agent framework that decomposes evaluation into three layers: a Planner Agent that designs strategies, Specialized Agents that perform role-specific assessments, and a Supervisor Agent that integrates their outputs. To support systematic evaluation, we also introduce a curated benchmark dataset of venture capital co-investments, featuring diverse firm attributes and ground-truth syndicates. Across 140 cases, PartnerMAS consistently outperforms single-agent and debate-based multi-agent baselines, achieving up to 10--15\% higher match rates. Analysis of agent reasoning shows that planners are most responsive to domain-informed prompts, specialists produce complementary feature coverage, and supervisors play an important role in aggregation. Our findings demonstrate that structured collaboration among LLM agents can generate more robust outcomes than scaling individual models, highlighting PartnerMAS as a promising framework for high-dimensional decision-making in data-rich domains.

Figures (6)

• Figure 1: An illustration of the research design. (A) Co-investors shortlisting. (B) High-dimensional feature selection. (C) Hierarchical MAS framework.
• Figure 2: Performance benchmark for Single Agent, Debate MAS, and PartnerMAS .
• Figure 3: Model performance comparison across settings.
• Figure 4: Agent performance grouped by Specialized Agent clusters. (A) Bubble chart of eight agent clusters, in which accuracy (x) vs importance rank (y, 1=highest), Bubble size = agents per cluster. (B) PartnerMAS accuracy vs normalized HHI (lower = more diverse) with trend line. Node size = cluster count. (C) PartnerMAS accuracy vs number of Specialized Agent clusters.
• Figure 5: Accuracy and feature focus of Specialized Agents under different backbones: (A) gpt-4.1-mini. (B) gpt-5-nano. (C) gpt-5-mini.