LLM-based Multi-Agent System for Simulating Strategic and Goal-Oriented Data Marketplaces

Jun Sashihara, Yukihisa Fujita, Kota Nakamura, Masahiro Kuwahara, Teruaki Hayashi · 2025

Read LLM-based Multi-Agent System for Simulating Strategic and Goal-Oriented Data Marketplaces by Jun Sashihara, Yukihisa Fujita, Kota Nakamura, Masahiro Kuwahara, Teruaki Hayashi (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-13233.txt.

Introduction

Jun Sashihara and 4 co-authors build a simulation of data marketplaces where buyers and sellers are language model agents. Real marketplaces trade datasets from third parties, yet how buyers, sellers, data, and rules shape each other is still poorly mapped. Each agent in the system holds clear goals and acts on its own.

Agents plan, search, buy, set prices, and update data while also guessing future demand. Tests compare purchase counts per dataset, per buyer, and repeat buys with patterns from real marketplaces. The language model system matched real trading patterns more closely than rule-based models and showed market trends forming over time.

Catalog record

CREATORJun Sashihara, Yukihisa Fujita, Kota Nakamura, Masahiro Kuwahara, Teruaki Hayashi
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE42 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-13233

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

LLM-based Multi-Agent System for Simulating Strategic and Goal-Oriented Data Marketplaces Jun Sashihara, Yukihisa Fujita, Kota Nakamura, Masahiro Kuwahara, Teruaki Hayashi Abstract Data marketplaces, which mediate the purchase and exchange of data from third parties, have attracted growing attention for reducing the cost and effort of data collection while enabling the trading of diverse datasets. However, a systematic understanding of the interactions between market participants, data, and regulations remains limited. To address this gap, we propose a Large Language Model-based Multi-Agent System (LLM-MAS) for data marketplaces. In our framework, buyer and seller agents powered by LLMs operate with explicit objectives and autonomously perform strategic actions, such as—planning, searching, purchasing, pricing, and updating data. These agents can reason about market dynamics, forecast future demand, and adapt their strategies accordingly. Unlike conventional model-based simulations, which are typically constrained to predefined rules, LLM-MAS supports broader and more adaptive behavior selection through natural language reasoning. We evaluated the framework via simulation experiments using three distribution-based metrics: (1) the number of purchases per dataset, (2) the number of purchases per buyer, and (3) the number of repeated purchases of the same dataset. The results demonstrate that LLM-MAS more faithfully reproduces trading patterns observed in real data marketplaces compared to traditional approaches, and further captures the emergence and evolution of market trends. I Introduction Data have emerged as a tradable economic resource, and data marketplaces that mediate the purchase and exchange of datasets from third parties have rapidly expanded [ 1 ] . These marketplaces streamline data collection that previously required substantial cost and effort, while also providing organizations and researchers with access to diverse, high-quality datasets. As a result, they are increasingly recognized as critical infrastructures that accelerate innovation based on data that were closed within individual organizations [ 2 ] . Despite this progress, our understanding of how interactions among market participants, data, and regulations shape market dynamics remains limited. Smooth and efficient data transactions require well-designed and robust data marketplaces [ 3 ] . The importance of simulation-based approaches to bridge the gap between market design theory and practice has been emphasized [ 4 ] . However, most existing simulations rely on model-based approaches with simple, predefined trading criteria. Such approaches are insufficient to capture essential real-world characteristics, including the heterogeneous objectives of participants, their strategic decision-making processes, and the emergence of domain-specific or temporal trends. Consequently, these models cannot adequately reproduce or analyze key phenomena such as price formation, long-term demand fluctuations, repeated data usage, and value updates. To address these challenges, we propose a Large Language Model-based Multi-Agent System (LLM-MAS) for simulating data marketplaces. Buyer and seller agents in this system are endowed with explicit objectives and can reason in natural language to take strategic actions such as planning, searching, purchasing, pricing, and updating data. This design enables agents to reproduce complex decision-making, forecast demand, and adapt strategies to evolving market conditions. In contrast to conventional rule-based simulations, our framework supports broader, adaptive, and goal-oriented behavior selection through natural language reasoning. Figure 1 shows the overall architecture of our proposed system. Based on these motivations, this study makes the following contributions: 1. Market-level simulation architecture: We design an LLM-MAS that captures both macro-level market structures and long-term dynamics such as demand cycles and trend emergence. 2. Agent-level behavioral modeling: We implement detailed buyer and seller models capable of goal-oriented actions and natural language reasoning, capturing diverse decision-making and interactions. 3. Data market evaluation metrics: We introduce real-world evaluation metrics that connect micro-level agent behavior with macro-level market outcomes, validated against transaction records from an actual data marketplace. Fig. 1 : Overall architecture of the proposed LLM-MAS for data marketplaces. The system consists of seller agents providing and updating data, and buyer agents pursuing goal-oriented acquisitions, a GoalGenerator for assigning analytical objectives, and a DataGenerator for producing metadata. Embedded metadata in a vector database supports cosine similarity-based search, reproducing the strategic behaviors, demand fluctuations, and trend emergence observed in real-world data marketplaces. These contributions collectively provide a structural perspective on data marketplace dynamics: how micro-level strategic interactions among heterogeneous agents give rise to macro-level phenomena. Such a perspective lays the foundation for future research on market and institutional design in which, simulation-based insights can inform policy-making, pricing mechanisms, and data management strategies. II Related Work Data marketplaces, which include data suppliers, buyers, and platform operators [ 5 ] , serve as a medium for the exchange and monetization of data across various organizations, gaining significant attention as a vital infrastructure in the data economy [ 6 , 7 ] . However, the range of phenomena that can be directly observed in these markets remains limited [ 8 ] . To address this limitation, Fernandez et al. [ 4 ] emphasized the importance of employing simulation-based approaches to study data markets, arguing that model-based simulations are effective tools for analyzing market structures and trading mechanisms. Building on this idea, prior research has created networks in which data are represented as nodes, examined their structural properties [ 9 ] , conducted a scenario analysis of data purchasing behaviors [ 10 ] , and proposed simulations in which buyer agents perform random walks on a data network to evaluate the impact of purchasing strategies [ 11 ] . However, these approaches face two main obstacles. First, model-based techniques rely extensively on established behavioral rules, making it difficult to understand complex or adaptive decision-making processes [ 12 ] . Second, while data transactions are generally motivated by clear individual goals, these methods assume that all agents are homogeneous, overlooking the various reasons for participating in the market [ 10 ] . Moreover, in contrast to financial markets where stylized facts and established simulation benchmarks are present [ 13 ] , data markets do not have standardized market features or evaluation criteria. This lack of standardization impedes the use of current simulation methods in the practical development of real-world markets. The advent of LLMs has presented novel opportunities to address these constraints. Owing to their sophisticated capabilities in natural language comprehension and reasoning, LLMs facilitate innovative simulation frameworks [ 14 ] . For example, Generative Agents [ 15 ] introduced a framework in which agents powered by LLMs interact and operate over extended periods. ReAct [ 16 ] merges reasoning with action generation, enabling agents to adapt plans dynamically and engage with their surroundings to acquire new information. These frameworks have demonstrated that LLM-based agents are highly proficient in executing complex tasks and exhibiting emergent behaviors [ 17 ] . Building on this foundation, recent research has utilized LLM-MAS in social simulations. For instance, TwinMarket [ 18 ] has shown its capability in examining emergent phenomena that connect individual investor behaviors at the micro-level to broader financial market trends. AgentSociety [ 19 ] incorporated psychological theories and a three-layer social environment into an LLM-MAS framework to investigate societal challenges such as political polarization and the implementation of universal basic income. Our research leverages recent advancements by adapting LLM-MAS to data marketplaces. We create a market setting in which LLM-based agents function as data buyers and sellers, capturing diverse motivations and complex dynamics that traditional model-based simulations cannot address. In addition, we derive market-specific attributes from real transaction data

Read the complete text