---
title: 'PartnerMAS: Hierarchical Multi-Agent Partner Selection'
url: https://www.emergentmind.com/topics/partnermas
type: topic
---

# PartnerMAS: Hierarchical Multi-Agent Partner Selection

PartnerMAS is a hierarchical large-language-model multi-agent framework for high-dimensional business partner selection, introduced for the concrete task of venture-capital co-investor shortlisting. It treats partner choice as a shortlist-generation problem over large candidate pools described by heterogeneous numerical, categorical, and textual features, and organizes reasoning into a Planner Agent, multiple Specialized Agents, and a Supervisor Agent. In the reported benchmark of 140 venture-capital cases, PartnerMAS consistently outperforms single-agent and debate-style multi-agent baselines, with the best reported configuration reaching a 70.89% match rate and the overall improvement summarized as up to 10–15% higher match rates [2509.24046].

## 1. Problem formulation and decision scope

PartnerMAS is defined for high-dimensional decision-making in which a candidate pool
\[
C = \{c_1,\dots,c_m\}
\]
must be evaluated under task context \(\mathcal{Q}\). Each candidate \(c_j\) is represented by a mixed-type feature vector
\[
x_j \in \mathbb{R}^n,
\]
with the full data matrix
\[
X \in \mathbb{R}^{m \times n}.
\]
Given the context, the system outputs a shortlist
\[
S = (s_1,\dots,s_k), \quad S \in C^k, \quad k = \lfloor m/3 \rfloor.
\]
The operational target is not final partnership formation after negotiation, but the shortlisting stage: reducing a broad pool to a smaller set of plausible partners [2509.24046].

The difficulty arises from the simultaneous presence of heterogeneous features, large candidate pools, many potentially relevant dimensions, and context-dependent relevance. The paper emphasizes numerical variables such as deal count, IPO count, degree centrality, and tie strength; categorical variables such as state, firm type, and year-quarter; and textual fields such as geographic preferences, industry preferences, and investment-stage preferences. The central claim is that a single reasoning trajectory is easily overloaded by this combination of scale and heterogeneity, producing weak consistency and limited robustness [2509.24046].

A common misconception is to interpret PartnerMAS as a generic partner-negotiation protocol. In the reported system, it is instead a structured shortlist generator for high-dimensional mixed-feature evaluation. Its immediate application is business partner selection, specifically VC co-investor shortlisting, although the paper presents this as a representative instance of a broader class of high-dimensional decision problems [2509.24046].

## 2. Hierarchical architecture and aggregation logic

PartnerMAS decomposes the task into three layers: a Planner Agent (PA), a set of Specialized Agents (SA\(_1,\dots,SA_N\)), and a Supervisor Agent (SPA). The hierarchy is motivated by the claim that these tasks require not merely more interaction, but explicit separation between strategy design, role-specific analysis, and final aggregation [2509.24046].

| Layer | Primary function | Main output |
|---|---|---|
| Planner Agent | Interpret case context and design evaluation strategy | Specialist configurations and strategic guidance |
| Specialized Agents | Evaluate the candidate pool from distinct perspectives | Ranked shortlists with alignment scores |
| Supervisor Agent | Integrate specialist outputs | Final ranked shortlist |

The Planner reads the lead investor profile, the investment target profile, feature names, and a sample of candidate profiles. It produces two objects: a concise strategic guidance paragraph and a JSON array of specialist configurations. Each specialized-agent configuration includes an agent name, role, abilities, and profile or guides. The planner is therefore a decomposition module rather than a scorer [2509.24046].

The specialists operate primarily independently rather than dialogically. Each acts from a particular perspective, performs an implicit feature-selection step guided by its assigned profile, evaluates the entire candidate pool, and returns a ranked shortlist \(S'_i\) together with alignment scores \(score_{ij} \in [1,10]\). Each output includes an `"evaluation_focus"`, an `"overall_rationale"`, and `"ranked_candidates"`. The paper states that each specialist returns exactly the top \(k'=\lceil m/3 \rceil\) candidates. Since the overall problem statement uses \(k=\lfloor m/3 \rfloor\), the notation is internally inconsistent between the global task definition and the specialist-output description [2509.24046].

The Supervisor aggregates specialist outputs in two stages. First, it performs consensus selection by counting how many specialists include each candidate:
\[
F_{1}(c_j) = \sum_{i=1}^{N} \mathbb{I}(c_j \in F'_i).
\]
Second, it resolves conflicts using weighted specialist importance:
\[
F_{2}(c_j) = \sum_{i=1}^{N} w_i \cdot \frac{1}{R_i(c_j)},
\]
where \(w_i\) is the importance weight of specialist \(i\) and \(R_i(c_j)\) is candidate \(c_j\)'s rank under that specialist. The paper also describes a prompt-level implementation in which the Supervisor sums the weights of all agents recommending a candidate and ranks by the resulting weighted score. It does not fully reconcile this prompt-based aggregation with the rank-sensitive formulation, but both descriptions assign the Supervisor a substantive conflict-resolution role rather than a purely clerical one [2509.24046].

This architecture is explicitly contrasted with debate-based MAS. Specialists in PartnerMAS do not critique one another. Their complementarity comes from differentiated profiles and differentiated feature attention, while synthesis is centralized in the Supervisor [2509.24046].

## 3. High-dimensional feature representation and benchmark construction

The benchmark is built from venture-capital co-investment data. Two commercial sources are used: LSEG Workspace for VC investment records from 1980 to 2024, and PitchBook for lead-VC identification and lead-firm enrichment. After restricting to U.S.-based companies, excluding solo investments, filtering for complete first-round information and syndicates with at least three investors, retaining only high- or very-high-confidence lead-VC matches, and keeping only cases with a single lead VC, the sample is reduced to 2,218 companies. Additional merging and manual matching then produce the final experimental dataset of 140 cases [2509.24046].

The candidate pool is contextually constructed by year, state, and industry. Pairwise lead-VC/candidate-VC observations are also created to represent prior co-investment experience and geographic distance. The resulting benchmark contains mixed structured data, including identifiers and labels, target-company attributes, candidate-VC attributes, activity and outcome measures, pairwise network variables, and geospatial attributes [2509.24046].

The feature inventory includes fields such as `companyindustrymajorgroup`, `companynation`, `companystate`, `firmtype`, `firmnation`, `firmstate`, `firmgeographypreference`, `firmindustrypreference`, `firminvestmentstagepreference`, `vcfirm_dealcount_20qtr`, `vcfirmIPOcount_20qtr`, `boncent`, `degree`, and `pair_tie_strength`. Numerical, categorical, and textual information are not embedded through a learned tabular encoder inside the reported decision pipeline. Instead, the system passes structured profiles to LLM agents through prompts, relying on in-context reasoning over candidate descriptions rather than task-specific supervised training [2509.24046].

Ground truth is historical co-investment syndicates. In the feature table, `real = 1` denotes that a candidate VC appears in the actual syndicate, and `real = 0` otherwise. Evaluation therefore compares the generated shortlist \(F\) with the true partner set \(G\) [2509.24046].

The benchmark has explicit constraints. It is comparatively small, restricted to U.S. venture capital, not publicly releasable because of licensing and confidentiality, and not accompanied by a formal fairness or bias audit. These limitations directly affect reproducibility and external validity [2509.24046].

## 4. Experimental protocol, match-rate metric, and reported performance

The evaluation metric is Match Rate:
\[
\text{Match Rate} = \frac{|F \cap G|}{|G|} \times 100\%.
\]
This is effectively recall over actual partners. The paper illustrates the metric with an example in which \(m=36\), the shortlist size is \(k=12\), and \(|G|=4\); if 3 true partners are found, the intended match rate is 75%, although the accompanying sentence contains a small inconsistency in the set-intersection notation [2509.24046].

The baseline families are Single Agent and Debate MAS. The Single Agent baseline has one LLM review all candidates and produce a ranked shortlist; for some efficiency comparisons, a four-run version is also used. The Debate MAS baseline uses three specialized agents and one supervisor, with critique and revision steps before final synthesis. All experiments use temperature \(=0\) [2509.24046].

| Setting | Configuration | Match Rate |
|---|---|---:|
| Best reported PartnerMAS | `gpt-4.1-mini` backbone with business-domain guidance | 70.89% |
| Best cited single-agent baseline | `gpt-5` (medium effort) | 61.50% |
| Best cited single-agent baseline | `gemini-2.5-pro` | 61.42% |
| Debate MAS example | `gpt-4.1-mini` | 60.19% |

The headline result is the 70.89% match rate for PartnerMAS with `gpt-4.1-mini` and business-domain guidance. The abstract summarizes the overall gain as up to 10–15% higher match rates. The body text gives concrete examples of smaller but still substantial margins. For instance, a debate-style system with `gpt-4.1-mini` reaches 60.19%, well below the best PartnerMAS configuration with the same backbone family [2509.24046].

Business-domain guidance improves both Single Agent and PartnerMAS, but more strongly for the hierarchical framework. For Single Agent, the gain is reported as 2–5% absolute. For PartnerMAS, gains can exceed 7%, with one reported jump from 62.55% to 69.03% [2509.24046].

Backbone assignment within the hierarchy also matters. In the cross-backbone table, the best entry is obtained when the Supervisor uses `gpt-4.1-mini`, even when Planner and Specialists use `gpt-4o-mini`, yielding \(69.03\% \pm 5.94\). This is consistent with the paper’s claim that aggregation quality is a major driver of final performance [2509.24046].

The efficiency analysis reports that PartnerMAS can be both more accurate and cheaper in token usage than larger single-model baselines. One example states that PartnerMAS with `gpt-4.1-mini` achieves over 70% accuracy while using fewer tokens than `gpt-5 (medium effort)`. The average system size in the efficiency comparison is 4.27 agents, and the empirical agent-count analysis suggests that the best performance occurs with a modest number of active agents, approximately 4–5 [2509.24046].

## 5. Internal behavioral findings

The paper’s behavioral analysis treats PartnerMAS not merely as a black-box benchmark winner but as a structured division of labor whose components exhibit distinct sensitivities. The clearest finding concerns the Planner. Logistic and linear regressions indicate that planner behavior is driven more by prompt guidance and model choice than by fine-grained case attributes. The odds ratio for deploying an Industry & Sector agent under business-domain guidance is 57.61 \((p<0.001)\), and the corresponding odds ratio for Risk & Compliance is 23.06 \((p<0.001)\). Model choice is likewise a strong predictor; for Risk & Compliance deployment, the reported effect size is 141.94 [2509.24046].

This means that the planner behaves strategically but not strongly case-adaptively in the narrow sense of being finely tuned to contextual metadata. The business-guided prompt explicitly asks the planner to consider dimensions such as network and collaboration history, industry fit, strategic alignment, financial factors, and geography, and the planner responds sharply to that prompt template [2509.24046].

Specialized agents are dynamically configured, but recurring specialist clusters emerge from post hoc clustering over 9,786 generated profiles across 14 experimental settings. The reported clusters include Risk & Compliance, Industry & Sector, Financial Capacity, Strategic Value, Partnership History, Network Connectivity, Geographic & Location, and Investment Stage. These specialists provide complementary feature coverage: geographic agents emphasize geography and location, industry agents emphasize firm industry preference, and network agents emphasize `pair_tie_strength`, `degree`, and `boncent` [2509.24046].

Specialist quality varies strongly by role and backbone. Reported examples include Risk & Compliance with `gpt-4.1-mini` at 83.3%, Investment Stage at 37.7%, and Partnership History under `gpt-5-mini` at 92.5%. The paper also notes that `gpt-5-nano` specialists are more diffuse and that no cluster exceeds 70% accuracy under that backbone. This variability is one reason the Supervisor becomes consequential rather than redundant [2509.24046].

The Supervisor’s prioritization decisions correlate with final correctness. For `gpt-5-nano`, assigning top importance to Partnership History, Industry & Sector, and Geographic & Location is reported as highly predictive of correct outcomes \((p<0.001)\). For `gpt-4.1-mini`, success is linked to prioritizing Investment Stage \((p<0.001)\), Strategic Value \((p<0.01)\), and Network Connectivity \((p<0.01)\). These results support the interpretation that aggregation is not a passive fusion step but an active ranking of expertise [2509.24046].

## 6. Limitations, interpretive boundaries, and related traditions

The principal limitations are explicit. The final dataset contains only 140 cases; the domain is restricted to U.S. venture capital; the experiments rely heavily on advanced proprietary GPT-family backbones; planning is prompt-sensitive; the Supervisor can become a bottleneck; no formal fairness or bias audit is reported; the data cannot be publicly released because of licensing and confidentiality; and hosted LLM updates may induce result drift over time even at temperature 0 [2509.24046].

These constraints delimit the strength of the paper’s broader claims. The reported results support PartnerMAS as an effective benchmarked architecture for mixed-feature VC co-investor shortlisting. They do not yet establish domain-independent superiority across all high-dimensional partner-selection settings. A plausible implication is that the framework’s strongest transferable contribution lies in its decomposition strategy—planner, independent specialists, weighted supervisor—rather than in any single prompt or benchmark [2509.24046].

PartnerMAS also belongs to a broader lineage of partner-selection research, but it differs sharply from earlier traditions. MAPSS, for example, addressed partner-and-service selection in service-oriented virtual organization breeding environments through competence descriptions, social requirements, performance criteria, and a genetic algorithm over VO variants [1111.5721]. In another line of work, distributed enterprise coordination was modeled through Virtual Enterprise Nodes combining Negotiator Agents and Planner Agents, with tier-level and network-level mediation for unresolved supply problems [0806.3031]. PartnerMAS differs from both by locating the main computational burden in LLM-based reasoning over mixed structured profiles rather than in explicit genetic search or supply-chain negotiation protocols.

The name itself is not fully standardized across literatures. In unrelated cloud computer-algebra literature, one paper states that “PartnerMAS” appears most likely to be a naming confusion or variant referring to MathPartner/Mathpar rather than a distinct platform [2204.11549]. In the present sense, however, PartnerMAS denotes the hierarchical LLM framework for business partner selection on high-dimensional features [2509.24046].

Source: https://www.emergentmind.com/topics/partnermas