Scsimulator Achieves Supply Chain Partner Selection Via LLM-Driven Multi-Agent Simulation

Supply chain optimisation presents a persistent challenge for businesses, particularly when selecting reliable and effective partners. Shenghan Gao, Junye Wang, and Junjie Xiong from ShanghaiTech University, alongside Yun Jiang, Yun Fang and Qifan Hu from Peking University, address this critical issue with their new visual analytics framework, SCSimulator. This innovative system moves beyond traditional, often biased, methods by leveraging large language models (LLMs) to drive multi-agent simulation, accurately modelling the complex interplay of competition and collaboration within supply chains. SCSimulator’s significance lies in its ability to visualise dynamic network structures and behaviours, offering transparent explanations of decision-making trade-offs through a combination of Chain-of-Thought reasoning and explainable AI , ultimately empowering users to explore strategic outcomes and improve partner selection processes.

LLM-driven simulation for supply chain partners offers enhanced

Scientists have developed SCSimulator, a novel visual analytics framework designed to revolutionise partner selection within complex supply chains. This breakthrough integrates Large Language Model (LLM)-driven Multi-Agent Simulation (MAS) with human-in-the-loop collaboration, addressing critical limitations in current supply chain optimisation techniques. The research tackles the challenge of optimising supply chains, particularly the crucial bottleneck of partner selection, which is shaped by both competitive and cooperative dynamics, a problem fundamentally rooted in multi-objective dynamic game theory. Traditional methods, often relying on mathematical simplifications and subjective managerial heuristics, frequently fail to capture the intricacies of real-world scenarios and introduce potential biases.

The team achieved a significant advancement by moving beyond fixed, uniform agent logic in MAS, leveraging recent progress in LLMs to represent complex supply chain requirements and hybrid game logic with greater fidelity. SCSimulator simulates the evolution of supply chains through adaptive network structures and enterprise behaviours, presenting these dynamics through easily interpretable visual interfaces. By combining Chain-of-Thought (CoT) reasoning with Explainable AI (XAI) techniques, the framework generates multi-faceted, transparent explanations of decision trade-offs, enabling users to understand the rationale behind simulated outcomes. This allows for iterative adjustment of simulation settings, aligning outcomes with user expectations and strategic priorities, and fostering a more informed decision-making process.

Experiments show that SCSimulator effectively models dynamic supply chain relationships while ensuring interpretability and balancing agent autonomy with expert control. Developed through iterative co-design with supply chain experts and industry managers, the system serves as a proof-of-concept, offering both methodological contributions and practical insights for future research on supply chain decision-making and interactive AI-driven analytics. The framework’s ability to visualise complex interactions and provide transparent explanations is a key innovation, addressing a critical gap in existing simulation tools. Usage scenarios and a user study demonstrate the system’s effectiveness and usability, validating its potential for real-world application. SCSimulator opens new avenues for exploring supply chain dynamics, enabling stakeholders to proactively assess risks, identify optimal partnerships, and enhance overall supply chain resilience. This work establishes a foundation for future research into interactive, AI-driven analytics that can empower organisations to navigate the complexities of modern supply chain management with greater confidence and efficiency.

LLM-Driven Supply Chain Simulation and Explanation offers enhanced

Scientists developed SCSimulator, a visual analytics framework integrating Large Language Model (LLM)-driven Multi-Agent Simulation (MAS) with human-in-the-loop collaboration to address partner selection challenges within supply chains. The study pioneered a system capable of simulating supply chain evolution through adaptive network structures and enterprise behaviours, presenting these dynamics via interpretable interfaces for enhanced understanding. Researchers engineered the framework to harness the reasoning capabilities of LLMs, specifically employing Chain-of-Thought (CoT) reasoning alongside Explainable AI (XAI) techniques to generate multi-faceted, transparent explanations of decision trade-offs. The core of SCSimulator lies in its ability to model complex supply chain requirements and hybrid game logic using LLMs, moving beyond traditional MAS approaches reliant on mathematical models and static agent logic.

Experiments employ an adaptive network structure, allowing the simulation to dynamically adjust relationships between enterprises as the scenario unfolds, reflecting real-world complexities. The system delivers a visual representation of these evolving relationships, enabling users to observe how changes in one area of the supply chain propagate through the network. Scientists meticulously designed the framework to balance agent autonomy with expert control, allowing users to iteratively adjust simulation settings and explore outcomes aligned with their expectations and strategic priorities. To ensure transparency and interpretability, the team integrated CoT reasoning, prompting the LLM to articulate its decision-making process step-by-step.

This reasoning is then augmented with XAI techniques, highlighting the key factors influencing each decision and providing users with a clear understanding of the underlying rationale. The study pioneered a co-design methodology, collaborating with supply chain experts and industry managers throughout the development process to ensure the framework’s relevance and usability. This iterative co-design process informed the development of interpretable interfaces and facilitated the creation of realistic simulation scenarios. Developed as a proof-of-concept, SCSimulator offers both methodological contributions and practical insights for future research on supply chain decision-making and interactive AI-driven analytics.

A user study and usage scenarios demonstrated the system’s effectiveness, confirming its ability to support informed decision-making in complex supply chain environments. The approach enables exploration of various ‘what-if’ scenarios, allowing stakeholders to assess the potential impact of different strategies and identify robust solutions. This innovative framework represents a significant advancement in visual analytics for supply chain management, offering a powerful tool for understanding and optimizing complex network dynamics.

SCSimulator visualises dynamic supply chain partner selection

Scientists have developed SCSimulator, a novel visual analytics framework integrating Large Language Model (LLM)-driven Multi-Agent Simulation (MAS) with human-in-the-loop collaboration for supply chain (SC) partner selection. The research team successfully created a system capable of simulating SC evolution through adaptive network structures and enterprise behaviours, visualised via interpretable interfaces. Experiments revealed that the framework addresses key challenges in modelling dynamic SC relationships, ensuring interpretability, and balancing agent autonomy with expert control. The team measured the system’s ability to rapidly configure scenarios, accommodating diverse inputs such as firm profiles, scenario descriptions, and simulation parameters, achieving flexibility in simulation construction.

SCSimulator’s agents demonstrate context-aware autonomous decision-making, independently analysing and responding to scenario-specific information, and the system supports heterogeneous agent behaviour, avoiding the homogenised decision-making common in traditional MAS. Results demonstrate that agents possess independent decision-making logic, distinguished from global rules, allowing for diverse decision-making styles within the simulation. Further tests prove the system’s capacity to reveal agent decision logic and its impact, combining LLM reasoning with quantitative attribution mechanisms. Scientists recorded that this combination provides clear, interpretable explanations of agent behaviours and their systemic impacts on the supply chain network.

The framework visualises supply chain interactions, offering scalable and intuitive visualisations that allow users to trace the evolution of network structures over time, providing fine-grained insights into interactions between agents. Measurements confirm that SCSimulator facilitates expert involvement, supporting adjustments to agent priorities, modification of decision criteria, and customisation of simulation interactions. The system records and traces simulation paths, enabling experts to explore alternative trajectories based on emerging interests, and providing a flexible, dynamic approach to simulation exploration.