You’re listening to “Integrated large language models and digital twin-based decision support in closed-loop supply chain operations: green steel transition case study,” by C.A. Le and colleagues. Published in 2026. ISSN: 0020-7543 (Print) 1366-588X (Online) Journal homepage: the linked source Integrated large language models and digital twin-based decision support in closed-loop supply chain operations: green steel transition case study Chi An Le, Devika Kannan, Kannan Govindan & Truong Thinh Nguyen To cite this article: Chi An Le, Devika Kannan, Kannan Govindan & Truong Thinh Nguyen (21 Jul 2026): Integrated large language models and digital twin-based decision support in closed-loop supply chain operations: green steel transition case study, International Journal of Production Research, DOI: 10.1080/00207543.2026.2702057 © 2026 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group. Published online: 21 Jul 2026. Submit your article to this journal Article views: 635 View related articles View Crossmark data Citing articles: 1 View citing articles Integrated large language models and digital twin-based decision support in closed-loop supply chain operations: green steel transition case study Chi An Le∗, Devika Kannan∗, Kannan Govindan∗ and Truong Thinh Nguyen Centre for Sustainable Operations and Resilient Supply Chains (CSORSC), School of Management, College of Business and Law, Adelaide University, Adelaide, Australia ABSTRACT. The green steel transition requires corporates to operate closed-loop supply chains (CLSCs) that involve operational decisions under heterogeneous data. Although decision models have shifted CLSC operations towards advanced analytics and optimisation, challenges remain in automatically interpreting natural-language requests, enabling data-driven reasoning, and communicating out-comes for actionable decision-making. Motivated by recent progress in large language models (LLMs), we propose an intelligent digital twin-based decision-support method (LMiDTSC) to auto-matically interpret, reason, and address CLSC operational tasks, and synchronise physical and digital CLSC resources. We develop a knowledge graph, a benchmark, and an accurate query-explanation consistency (AQEC) metric to operate the method and evaluate its effectiveness in supporting deci-sions on steel distribution and scrap collection. Across focal operational tasks, LMiDTSC outperforms baselines (e.g. GPT-5.2) by 82.73% in AQEC, while its fine-tune outperforms backbones using pro-prietary models by 93.17% in AQEC. This study advances decision support systems and AI-enabled operations research by integrating LLMs and digital twins for computerised decisions in CLSC oper-ations. This study suggests supply chain practitioners should consider high-performing LLMs for complex operational tasks and exercise caution in scaling in-context learning prompts. Finally, policymakers can accelerate the implementation of such integrated decision support systems by promoting industrial symbiosis and workforce training. 1. Introduction. The steel supply chain plays an important role in mod-ern economic activity, from large-scale infrastructure to consumer goods. Yet the sector faces a pressing need for the decarbonisation challenge. The global steel value chain accounts for a substantial share of greenhouse gas emis-sions and energy consumption, creating strong regulatory and market pres-sure to transition to green steel. Achieving this transition requires cleaner production pathways and more effec-tive operational decision-making across procurement, distribution, and end-of-life recovery, especially where reverse flows of scrap play a central role in lowering the carbon footprint of steel products throughout their lifecycle. ARTICLE HISTORY Received 3 November 2025 Accepted 16 June 2026 Green steel; decision support system; large language model; digital twin; graph-based retrieval-augmented generation; closed-loop supply chain A major problem is that many emissions associated with steel supply chains arise beyond direct plant oper-ations. This challenge is formally structured in carbon accounting through three emission scopes, where Scope 1 refers to direct emissions from owned or controlled sources, Scope 2 to indirect emissions from purchased energy, and Scope 3 to other indirect emissions occurring across the upstream and downstream value chain. For steel, Scope 3 can be heav-ily influenced by logistics, procurement choices, inven-tory positioning, and end-of-life recovery. Consequently, implementing closed-loop supply chain (CLSC) operations, such as increasing scrap collection and recycling for secondary steelmaking, and coordinating forward and reverse distribution, can materially affect both cost and sustainability performance. However, translating these strategic roadmaps into day-to-day operational actions remains difficult because CLSC decisions are typically dynamic and multi-entity. From an operational perspec-tive, CLSC operations must manage uncertain demand and returns, coordinate real-time and interpretable deci-sions, and handle the complexity of multi-layered net-works. Decision support systems for such settings often rely on optimisation and simulation techniques rooted in linear and non-linear programming. While powerful, these approaches impose a high formulation burden, where supply chain practi-tioners must translate managerial questions into math-ematical models, specify constraints, and interpret solver outputs into actionable guidance. In practice, this gap between natural language decision needs and model inputs and outputs can limit adoption, particularly when decisions require real-time visibility, optimisation, what-if analysis, and explanations. Digital twin technology offers a promising founda-tion to reduce this gap. Digital twin provides continu-ously updated virtual representations of physical systems, and prior literature has explored their potential to sup-port real-time monitoring, simulation, and optimisation in operations and supply chain management, including CLSC operations. In typical imple-mentations, however, the decision intelligence of digital twin systems continues to rely on mathematical solvers and customised analytical workflows. As a result, sup-ply chain practitioners must still possess sufficient tech-nical expertise to translate ad hoc managerial requests into executable model queries, update assumptions and parameters, and interpret solver outputs into narratives to support decision-making processes. This limits the usability of digital twins when operational interactions are primarily conversational, exploratory, and iterative. Recent progress in large language models offers a com-plementary pathway to address this limitation. These models can interpret natural-language queries, gener-ate programming code, and produce contextual expla-nations, making them well-suited to bridging the gap between conversational decision-making needs and for-mal model interactions. Recent literature has docu-mented significant contributions of large language mod-els in operations and supply chain management (Chen, Kirshner, et al. 2025; Fosso Wamba et al. 2024; Jack-son et al. 2024; Kirshner et al. 2025). Beyond supply chain contexts, emerging work has begun to integrate large language models into digital twin systems to add a cognitive reasoning layer over both structured and unstructured data, enabling more natural interactions and adaptive decision support in diverse domains. Exam-ples include large language model-assisted digital twin workflows for virtual clinical trials, personalised alert generation for infrastructure mainte-nance, and adaptive automation and human-robot collaboration. These studies illustrate the potential of large language models to reduce interaction barriers and enhance reasoning and explanation in digital twin-based decision support systems. However, large language models are sometimes not reliable decision engines in domain-specific operational settings because they can generate plausible but ungrounded content, known as hallucinations. Effective deployment, therefore, requires grounding large language model reasoning in enterprise data and constraints, together with reliable execution layers. For CLSC operational decision support, knowl-edge graphs are natural substrates for grounding because supply chains are inherently network-structured, and graph-based query languages can express multi-hop rela-tionships, feasibility constraints, and operational paths. Despite this conceptual fit, the literature still lacks a tai-lored methodology showing how large language mod-els and knowledge graphs can be integrated with dig-ital twin-based supply chain decision support systems to deliver graph-grounded reasoning, interactive and explainable decision support, and synchronisation with structured and unstructured operational data, particu-larly for CLSC operations. Therefore, it is necessary to develop an intelligent dig-ital twin-based decision support approach that can inter-pret users’ requests expressed in natural language, trans-late them into reasoning steps over the knowledge graph, and return grounded explanations suitable for oper-ational decision-making, without imposing modelling and solver-configuration burdens typically associated with linear programming workflows. Motivated by these research gaps, this study proposes a novel large language model-augmented intelligent digital twin-based decision support method for CLSC operations (LMiDTSC). The research scope is to implement and evaluate the devel-oped method in the context of steel CLSC operations, thereby supporting decision-making that underpins the transition towards green steel value chains. The proposed method represents the CLSC as a supply chain knowledge graph and adopts the graph-based retrieval-augmented generation for multi-hop reasoning. It leverages the text generation capability of large language models to reason over the knowledge graph, while drawing on the digital twin concept to enable real-time visualisation, optimisa-tion, and cyber-physical synchronisation through com-mon operational tasks such as supply chain visibility, distribution network optimisation, and what-if analysis. As a result, supply chain practitioners are not required to manually formulate optimisation models and counterfac-tual scenarios for what-if analysis, configure optimisation solvers, or train task-specific artificial intelligence (AI) models to obtain actionable decision outputs. Operationally, the LMiDTSC implements a multi-step reasoning process. Given a user request in natural language, the proposed method first computes its text embeddings to retrieve semantically similar exemplars for an in-context learning prompt. The large language model then generates a graph database query to obtain authoritative results. A subgraph is extracted to support digital twin-based visualisation of relevant supply chain entities and relationships, and a follow-up large language model step produces a grounded managerial explanation. The method further supports cyber-physical synchroni-sation through structured enterprise information system updates and adaptive knowledge updates for facilitat-ing continuous learning through the integration of novel operational tasks into the vector database. Conceptually, the LMiDTSC constitutes a domain-specific and human-centric decision support system for computerised deci-sions in operating a CLSC. It is designed to be usable for supply chain practitioners without a computer sci-ence background, while maintaining traceability between what is executed within the digital twin-based decision support system and what is communicated to users. In doing so, the LMiDTSC provides an empirical valida-tion and a practical pathway towards intelligent digital twin systems in supply chain management, where the digital twin is not only a real-time optimisation-simulation platform, but also an interactive reasoning system supporting decision-making processes. In summary, the main contributions of this paper are fourfold as follows. • This study proposes LMiDTSC, a large language model-augmented intelligent digital twin-based deci-sion support method that interprets supply chain operational requests in natural language, performs multi-hop reasoning over a supply chain knowledge graph, explains results, supports interactive visualisa-tion for focal operational tasks (i.e. supply chain visi-bility, distribution network optimisation, and what-if analysis), and enables cyber-physical synchronisation. • This study constructs a domain-specific knowledge graph and a multi-task benchmark specialised in a case study on the steel product and scrap distribution network, and a tailored evaluation metric, accurate query-explanation consistency (AQEC), to evaluate the effectiveness of the proposed method. • This study conducts intensive experiments that eval-uate the effectiveness of the proposed method across the focal operational tasks, with participation of large language models from the GPT and Llama families and reporting AQEC, together with common evalua-tion metrics in natural language processing (i.e. BLEU, ROUGE, BERTScore, and LLM-Judge). • This study examines the impact of fine-tuning on the proposed method, taking into account open-source and lightweight large language models, and evaluating performance against the incorporation of state-of-the-art and proprietary models using the aforementioned evaluation metrics. The remainder of this study is organised as follows. Section 2 reviews digital twin-based decision support for CLSC operations, large language model-enabled supply chain analytics and optimisation, and approaches for syn-ergising large language models and digital twins. Section 3 details the proposed empirical method. Section 4 describes the empirical setting, followed by the empirical findings and discussion in Section 5. Section 6 discusses implications. Finally, Section 7 concludes this study, indi-cates limitations, and outlines future work directions. 2. Literature review. This section summarises current approaches to digital twin-based decision support in CLSC operations, appli-cations of large language models in supply chain analytics and optimisation, and approaches to synergising large language models and digital twins. 2.1. Digital twin-driven decision support for CLSC. operations Decision support is crucial for enhancing the efficiency of CLSC operations. Over the last decade, prior works have capitalised on linear and non-linear programming and game theory approaches to develop many decision models for CLSC operations, such as circular procure-ment, recycling and remanufacturing, waste collection, carbon reduc-tion, and stakeholder collabora-tion. However, several challenges remain in conventional decision support sys-tems, including difficulty in precisely capturing decision-maker inputs expressed in natural language, complexity of modelling multiple supply chain players with one-to-many or many-to-many relationships, and limitations in han-dling unstructured data. Data-driven approaches, such as digital twins, have transformed decision support systems for CLSC operations by incorporating real-time data and virtual representations of physical objects, thereby enabling advanced modelling, simulation, and optimi-sation across forward and reverse supply chain func-tions. For instance, Guidani, Ronzoni, and Accorsi (2024) developed a digital twin-based traceability architecture that virtualises the agri-food supply chain to support decision-making processes from farming to retailing. Using a digital twin, Maheshwari et al. (2023) developed a real-time planning and monitoring approach considering industrial symbiosis to support decision-making in food supply chains based on agent-based simulation. Wang and Wang (2019) proposed a digital twin-driven recy-cling platform for making decisions in waste electrical and electronic equipment returns. Cao et al. (2023) used an optimisation method to develop digital twin-driven decision models for a medical waste clearance scheme. Kroell et al. (2024) applied a digital twin to enhance waste sorting and material recycling decisions in design and plant-level operations. Additionally, prior works paid sig-nificant attention to digital twin-driven decision support for enhancing disassembly productivity and safety, particularly human-robot collaborative disassembly for hazardous waste and end-of-life lithium batteries. The aforementioned works on digital twin-based decision support in CLSC operations primarily build on structured data with limited capability to incorpo-rate unstructured data. Thus, these approaches remain constrained in capturing practitioners’ intent expressed in natural language, reasoning over complex CLSC interactions at scale, and providing a cognitive layer that supports automated interpreting and addressing for operational decisions. Moreover, these works are often restricted in fully operationalising key digital twin concepts: visualisation, virtualisation, simulation-optimisation, and cyber-physical synchronisation. Particularly, an automa-tion remains challenging when moving from interpret-ing natural language requests to generating actionable insights that support comprehensive real-time visualisa-tion, optimisation, and synchronisation between phys-ical and digital supply chain processes. To address these challenges, this study uses steel supply chain enti-ties to model and construct a heterogeneous CLSC knowledge graph. This knowledge graph captures com-plex, dynamic interactions among supply chain enti-ties, thereby supporting real-time decision-making for steel CLSC operations. Building on this foundation, this study combines large language models, supply chain knowledge graph, and graph-based retrieval-augmented generation into the proposed intelligent digital twin-based decision support method that per-forms an interpretation of user requests in natural lan-guage (unstructured data), reasoning and retrieval of information embedded in the CLSC knowledge graph. Collectively, this proposed method supports corporate decision-making by providing interpretable answers for supply chain visibility, linear programming-based opti-misation, what-if analysis, and cyber-physical synchro-nisation in CLSC operations in real time. To the best of the authors’ knowledge, this study constitutes the first attempt to integrate a large language model and digi-tal twin for decision support in CLSC operations with a focus on the steel industry. 2.2. Large language models in supply chain. analytics and optimisation Recent progress in large language models has demon-strated significant potential for supply chain analytics and optimisation, thereby enhancing decision-making processes across supply chain functions (Chen, Kirsh-ner, et al. 2025; Fosso Wamba et al. 2024; Jackson et al. 2024). For instance, Li et al. (2023) developed a large language model-based framework to communicate what-if analysis from optimisation models’ outcomes with-out substantial modelling efforts of supply chain prac-titioners. Large language model-based approaches were employed to quantify firm-level supply chain risk expo-sure using transcripts from corporate site visits and to reveal significant risk reduction effects of digital innovation. Further-more, Phung, Huynh, and Khai (2025) addressed mis-information impacts on supply chain sustainability by leveraging large language models to detect fake news. Wang et al. (2025) proposed a large language model-enhanced text-mining workflow that supports financial service providers in supply chain finance analysis. Kir-shner et al. (2025) investigated large language model agents in autonomous supply chain contract negotia-tions, revealing that such agents demonstrate a greater inclination towards reaching agreements. Li, Ko, and Ameri (2025) developed a methodology that integrates large language models with knowledge graphs for sup-plier discovery to achieve high accuracy in supplier iden-tification tasks. Lastly, Gezdur and Bhattacharjya (2025) demonstrated how large language models can enhance supply chain employee training by reducing 25% training costs and providing personalised learning experiences. Among current research works, the literature has paid limited attention to applied studies using large language models to support reverse logistics and CLSC operations. While Li et al. (2023) and Li, Ko, and Ameri (2025) demonstrate the potential of large language models in reasoning about and interpreting optimisation outcomes in settings such as shipping planning and supplier dis-covery, their focus remains largely on forward procure-ment and logistics. The graph-based retrieval-augmented generation-driven supplier discovery framework of Li, Ko, and Ameri (2025) primarily extracts relationship pat-terns from knowledge graphs. However, it is structurally simpler than closed-loop operational decisions, which require reasoning across heterogeneous entities, reverse flows, and multi-hop network interactions. Although many works have contributed to information extrac-tion embedded in supply chain and industrial knowledge graphs through a ‘quintuplets’ initiative, or named entity and intent recognitions, they typically focus on extracting entities and relationships for downstream data analytics, rather than leveraging them to reasoning for complex opera-tional tasks, such as linear programming-based optimi-sation or what-if analysis. More broadly, there remains substantial scope for future research to clarify the oper-ational value of generative AI (GenAI) together with digital twins in driving sustainable supply chain manage-ment and converging physical and digital organisational resources. In response, this study employs large language models and the concept of digital twin to develop an intelligent decision support method that addresses supply chain visibility, linear programming-based optimisation, and what-if analysis within a steel product and scrap dis-tribution network. Furthermore, the proposed method supports supply chain practitioners in cyber-physical synchronisation between enterprise information systems (physical supply chain processes) and supply chain digital twins. 2.3. Approaches for synergising large language. models and digital twins Large language models can promote intelligent digital twin-based decision support in dynamic network envi-ronments by adding reasoning layers over unstructured data. The adoption of large language models in digital twin-based deci-sion support systems has been fuelled by the increasing availability of both structured and unstructured data and recent methodological advances. For instance, TWIN-GPT proposes a workflow that lever-ages ChatGPT to generate personalised digital twins tailored for virtual clinical trial simulation when lim-ited electronic healthcare records. TailorAlert provides a large language model-based alert gen-eration system for digital twin-based road infrastructure management. The system achieves an accuracy of 100% and a role-content match rate of 82%, thereby enhanc-ing decision-making in road maintenance and improv-ing infrastructure reliability and public safety. ChronoL-lama (Wang, Zhang, et al. 2025) introduces a fine-tuning framework for simulation-based multi-physics digital twin generation. Xia, Jazdi, and Weyrich (2025) present an architecture integrating a large language model and digital twin for an adaptive automation system, validated through a case study on robotic automation control and process simulation. Amad, Astorga, and van der Schaar (2025) treat digital twins as models of real-world sys-tems and frame them as an in-context learning problem. The authors propose CALM-DT, a simulation approach that models system dynamics across state-action pairs to minimise the generation error at each simulation state. Qin and Yang (2026) present an adaptive framework, DT-CSUniTime, that combines digital twin and large language models to address online lithium-ion battery capacity estimation, enabling more accurate capture of battery degradation features. Sun et al. (2024) propose a large language model-empowered digital twin multi-agent framework for decision-making in manufactur-ing resource scheduling and diagnostic decision-making using time-series data. Zhang et al. (2024) investigate the role of large language models in providing natural-language explanations for autonomous decisions in a dig-ital twin system of a drone fleet in smart farming. Zhang et al. (2025) propose a question-and-answer method that integrates a large language model-driven algorithm to realise geospatial digital twins, which represent dynamic changes to geographic objects and reduce professional barriers between knowledge graphs and users. Existing methods that integrate large language models into digital twin systems for intelligent network opti-misation are often developed as conceptual frameworks or proof-of-concept demonstrations. As a result, these approaches provide limited evidence on task-driven opti-misation performance and user-experience requirements for practical deployment. Many studies do not explicitly evaluate comprehensive decision support performance for complex operational tasks that require both formal optimisation and what-if analysis, together with inter-active visualisation. For domain-specific knowledge and tasks, hallucinations in large language models can prop-agate bias into digital twin outputs and dilute decision accuracy. This highlights the need for reliable knowl-edge organisation and information execution layers. In addition, performance comparisons between fine-tuned and non-fine-tuned large language models within digi-tal twins are also limited, leaving it unclear when fine-tuning offers consistent benefits over in-context learning prompting. Therefore, the operational value of incorpo-rating large language models into digital twin-based deci-sion support methods for operations and supply chain management remains largely unclear. In addition, this study is not aware of previous work that integrates large language models and digital twins to provide decision support in sustainable supply chain management, espe-cially for CLSC operational decisions. 2.4. Summary of research gaps. Despite significant methodological advances in integrat-ing large language models into digital twin-based deci-sion support methods discussed above, several gaps remain, especially with respect to operations and supply chain management, and production research. • Regarding CLSC operational decision support, exist-ing digital twin-driven approaches primarily rely on structured data and are often limited in their ability to capture natural-language user intent, reason over heterogeneous knowledge graphs, and provide a cog-nitive layer that supports automated optimisation and what-if analysis alongside visualisation and synchro-nisation. Accordingly, there remains a gap in developing an intelligent digital twin-based decision sup-port system that operationalises comprehensive digi-tal twin concepts, from interpreting requests to exe-cuting analytics, optimisation-simulation, and cyber-physical synchronisation in the CLSC operational set-ting in real time. • In supply chain analytics and optimisation, prior studies have shown large language models’ val-ues in interpreting outcomes, extracting insights from text, and supporting forward supply chain functions, yet applied evidence in reverse supply chains and closed-loop operations remains lim-ited. More importantly, existing approaches often focus on simplified tasks (e.g. supplier discovery) or result communication, rather than supporting operational decision-making that requires interac-tive optimisation-simulation, scalable reasoning, and robust handling of semantic diversity in user requests without incurring high computational cost. In addi-tion, modelling and case study research on the inte-gration of large language models and digital twins to provide tailored supply chain operational decision support for closed-loop implementation remains rare. This study aims to address these gaps by proposing LMiDTSC, a large language model-augmented intelligent digital twin-based decision support method for CLSC operations that integrates a domain-specific knowledge graph, graph-based retrieval-augmented generation, and query-based reasoning to support supply chain visibility, distribution network optimisation, and what-if analysis, real-time visualisation, and cyber-physical synchronisa-tion through natural-language interaction. By formulat-ing optimisation models and what-if analysis in the form of graph database queries and enabling interpretable answer generation, interactive visualisation, and knowl-edge graph updates for synchronisation, the proposed method empirically clarifies the value of synergising large language models, knowledge graphs, and digital twins for decision support in steel CLSC operations. 3. Empirical method. This section presents the problem definition and pro-posed method. The proposed method consists of four modules: knowledge graph-driven task interpretation and query grounding, knowledge graph-enabled multi-hop reasoning for information retrieval, explainable deci-sion synthesis and natural-language interaction, and dig-ital twin synchronisation and adaptive knowledge graph update (see Figure 1). 3.1. Problem definition. This study examines whether large language models can be incorporated to enhance the reasoning and automa-tion of digital twin-based decision support systems in CLSC operations. The goal is to develop a large language model-augmented digital twin decision support method that retrieves information from a supply chain knowledge graph to answer a user query expressed in natural lan-guage related to focal operational tasks (i.e. supply chain visibility, distribution network optimisation, and what-if analysis). This work focuses on a forward and reverse dis-tribution network of steel products and scrap in the steel industry in Vietnam. The case study is particularly well-suited to the scope of this study for two reasons. First, Vietnam has been one of the world’s leading steel produc-ers and exporters, facing challenges in maintaining global competitiveness while achieving alignment with national decarbonisation targets. Second, Vietnam’s steel value chain encompasses vast and intricate networks of stakeholders across pro-duction, distribution, and recycling processes. This char-acteristic provides a heterogeneous data source for con-structing a dynamic supply chain knowledge graph and evaluating the proposed method’s effectiveness in supporting decision-making in real-world CLSC operations. Therefore, the proposed method would not only help Vietnamese corporations effectively operate CLSCs tar-geting domestic sustainable steel value chains, but also provide a reference case for global steel corporations seeking to advance their green steel transitions. Specifically, the proposed method leverages large lan-guage models to generate graph database queries for multi-hop reasoning and information retrieval from the supply chain knowledge graph, and explain the results to address focal operational tasks. Let Q denote the user query for a focal operational task expressed in natu-ral language, P (Q) denote the task-grounded prompt, S denote the schema of the knowledge graph G = (V, E, μ, δ), where V is a set of entities, E is a set of relationships, μ is a set of entity attributes, and δ is a set of relationship attributes. Let LLLM denote a large language model, ˆC and ˆC GT denote generated and ground-truth graph database queries, respectively. Let ˆY and ˆY GT denote results obtained from executing ˆC and ˆC GT, respectively. Let ˆA and ˆAGT denote explanations for ˆY and ˆY GT, respectively. The aim of the proposed method is to maximise the quantitative and semantic similarities sim(ˆY, ˆY GT) and sim(ˆA, ˆAGT), where ˆC = LLLM (P (Q), S), ˆA = LLLM (Pans(Q, ˆY)), and sim(·, ·) are the evaluation metrics (see Section 4.2). Then, it aims to produce a subgraph ˆG Q = (VQ, EQ, μQ, δQ) where VQ ⊆ V, EQ ⊆ E, μQ ⊆ μ, δQ ⊆ δ for digital twin-based visualisation relevant to answering the query Q. Finally, the proposed method supports informa-tion updates derived from operational changes through a synchronisation function (i.e. create, read, update, or delete entities, relationships and attributes) Gt+1 = sync(Gt, 1t), where Gt = (Vt, Et, μt, δt) is the supply chain knowledge graph state at a timestamp t and 1t can be induced by user-initiated actions or changes of the external environment. 3.2. Knowledge graph-driven task interpretation. and query grounding This module interprets a user’s request in natural lan-guage and generates a grounded, executable graph query that is aligned with knowledge representation through semantic retrieval for in-context learning, and schema-constrained grounding to reduce hallucinations of large language models. The knowledge representation here includes a supply chain knowledge graph and its data schema. A text embedding generator f (·) encodes the user request Q into a vector embedding q = f (Q) ∈ Rd. Given a vector database B = {bi, Qi, ˆC GTi }N i=1, where bi is the vector embedding of exemplar request Qi and ˆC GTi is its ground-truth graph database query, the module com-putes cosine similarity in Equation and retrieves the top - k most similar exemplars in Equation: cos(q, bi) = qT bi qbi cos{(q, bi)}N (︁)︁ Rk(Q) = top-k i=1 The retrieved exemplar set Rk(Q) is then put into an in-context learning block consisting of pairs of questions and graph database queries (Qi, ˆC GTi). This in-context learning (ICL) block is concatenated with a predefined system prompt template P0 (see Appendix 1), yielding a task-grounded prompt P (Q) = P0 ⊕ ICL(Rk(Q)), where ⊕ denotes prompt concatenation. To comply with the supply chain graph data schema, P0 specifies an authoritative graph schema S that defines node labels, relationship types, and properties. In other words, a gen-erated graph database query ˆC only uses node labels, relationship types, and properties permitted by S. For-mally, the module maps Q to a focal operational task using the prompt instructions and retrieved exemplars. Finally, the module supports selecting a large lan-guage model LLLM from a candidate pool based on the estimated request difficulty and the in-context learning context size. In practice, this selection follows a heuris-tic routing rule, where lightweight models are preferred for simpler visibility queries, and high-performing mod-els for optimisation and what-if analysis reasoning. The module output (P (Q), LLLM, S) is then consumed to produce a graph database query and trigger multi-hop reasoning over the knowledge graph. 3.3. Knowledge graph-enabled multi-hop. reasoning for information retrieval This module realises multi-hop reasoning over a sup-ply chain knowledge graph for information retrieval to answer the focal operational tasks. It takes a package (P (Q), LLLM, S) to output a graph database query and its execution results, and a subgraph that serves as a dig-ital twin artefact for visibility and synchronisation. Con-ditioned on the prompt P (Q), the selected model LLLM generates a graph database query ˆC = LLLM (P (Q)) con-strained by the schema S. For distribution network opti-misation requests, the LMiDTSC resolves distribution planning directly on the knowledge graph by generating a graph database query that implements a graph-native greedy optimisation algorithm. This study describes a unified steel distribution optimisation model that covers forward and reverse distribution flows through role map-ping on the same property knowledge graph. The model objective is to meet sink demand with minimum total shipping and handling cost, where the one-stage setting is restricted to a single intermediate layer when the inter-mediate’s inventory is sufficient (inventory-push mode), and the two-stage setting allows replenishment decisions when the intermediate’s inventory is insufficient (sourc-ing mode). Importantly, the proposed method does not forgo the state-of-the-art combinatorial optimisation technology or focus on advancing solution methodol-ogy in operations research. Instead, it leverages a graph database query-based, greedy algorithm-inspired solu-tion method to quantitatively answer optimisation and what-if analysis tasks. Let N denote a set of sources, M denote a set of inter-mediates, K denote a set of sinks, and P denote a set of products: steel products and steel scrap. Let Dk,p denote the demand of sink k ∈ K for product p ∈ P, Im,p denote the initial inventory of product p ∈ P at intermediate m ∈ M, and Cn,p denote the supply capacity of source n ∈ N for product p ∈ P. Let cn,m,p denote the shipping cost from source n to intermediate m for product p ∈ P, cm,k,p denote the shipping cost from intermediate m to sink k, and let hm denote the handling cost at intermedi-ate m ∈ M (e.g. handling products and sorting or storing scrap). Decision variables are xm,k,p, the quantity shipped from intermediate m ∈ M to sink k ∈ K, and yn,m,p, the quantity shipped from source n ∈ N to intermediate m ∈ M. The model objective of the optimisation problem is to minimise the total cost in Equation. ∑︂ ∑︂ ∑︂ (cm,k,p + hm)xm,k,p Minimise Z = m∈M p∈P k∈K ∑︂ ∑︂ ∑︂ cn,m,pyn,m,p + n∈N m∈M p∈P The proposed optimisation model contains the follow-ing constraints: ∑︂ xm,k,p = Dk,p, ∀k ∈ K, ∀p ∈ P (3.i) m∈M ∑︂ ∑︂ xm,k,p ≤ Im,p + yn,m,p, ∀m ∈ M, ∀p ∈ P n∈N k∈K (3.ii) ∑︂ yn,m,p ≤ Cn,p, ∀n ∈ N, ∀p ∈ P (3.iii) m∈M xm,k,p, yn,m,p ≥ 0 (3.iv) The proposed optimisation model becomes four opti-misation problem settings by mapping (K, M, N, P) and the cost terms to the corresponding entities in the supply chain knowledge graph. • Forward one-stage distribution: K = Customers, M = Warehouses, N = ∅ (y ≡ 0), P = Steel. • Forward two-stage distribution: K = Customers, M = Warehouses, N = Suppliers, P = Steel. • Reverse one-stage distribution: K = Suppliers, M = Return Centres, N = ∅ (y ≡ 0), P = Scrap. • Reverse two-stage distribution: K = Suppliers, M = Return Centres, N = Customers, P = Scrap. This work proposes Algorithm 1 as a graph-native greedy optimisation algorithm within the proposed LMiDTSC method that prioritises direct fulfilment from interme-diate inventory and then extends to two-stage replenish-ment through eligible sourcing when the sink demand remains unmet. This design allows optimisation to be executed directly on the supply chain graph knowledge through graph database queries, while preserving inter-pretability through allocation records and cost break-down. Algorithm 1 is a heuristic that prioritises direct fulfil-ment from existing intermediate inventory to avoid the additional source-to-intermediate shipping cost incurred by two-stage replenishment, falling back to two-stage sourcing only when intermediate inventory is insuffi-cient. It does not guarantee global optimality of the inte-ger programme in Equation, but its greedy structure preserves interpretability through transparent allocation records and supports direct execution as a graph database query, which is the property that enables real-time opti-misation and natural-language traceability within the proposed LMiDTSC method. For what-if analysis requests, the prompt requires the large language model to modify the graph database query generated for the prior optimisation request. Formally, let 1 denote the counterfactual constraints in the new request (e.g. demand fluctuation, capacity reduction, or cost change). The intended behaviour is regenerated as ˆC cf = LLLM (P (Q), ˆC base, 1). Overall, the execution of ˆC base or ˆC cf on the graph database runtime returns a triple (ˆY, σ, l), where ˆY is the query’s execution result, σ ∈{0, 1} indicates execution succeeds, and l is the run-time error trace when σ = 0. To improve robustness against syntax-level failures (e.g. property name mis-matches or malformed graph traversals), the error trace l(t) is concatenated with the original prompt and fed back to the large language model to regenerate the query (t+1) = LLLM (P (Q) ⊕ l(t)), t = 0..., T − 1. The as ˆC loop terminates when σ = 1 (successful execution) or when the attempt budget T is exhausted. To support dig-ital twin-based visualisation, the module extracts a sub-graph ˆG Q = (VQ, EQ, μQ, δQ) corresponding to the user query, obtained by executing a visualisation-oriented variant of the generated query (see Appendix 2). 3.4. Explainable decision synthesis and. natural-language interaction This module translates a graph query’s output into an explainable decision response, where the explanation must be grounded in the graph query’s output to min-imise hallucination and to ensure decision reliability. It takes ˆC, (ˆY, σ, l), and ˆG Q as inputs and produces an explanation ˆA in natural language. If σ = 1, the LMiDTSC method constructs a follow-up query that requires the model LLLM to generate an explanation grounded exclusively in ˆY. The follow-up prompt is denoted by Pans(Q, ˆY) (see Appendix 1) and the final response is ˆA = LLLM (Pans(Q, ˆY)). As a result, the generation of ˆA is constrained by the executed data, reducing the likelihood that a large language model introduces unsupported facts. If the self-repair loop in Section 3.3 exhausts its T attempts without achieving σ = 1, the persistent error trace l is forwarded to this module, and a readable failure message is produced as ˆAerr = LLLM (Pans(Q, l)). This design supports trans-parent interaction by communicating unresolved query failure to the user rather than fabricating an answer. Algorithm 1: Graph-native greedy optimisation for product distribution Although the same grounding mechanism is applied across focal tasks, the decision semantics of ˆA dif-fer by task. For supply chain visibility, ˆA summarises descriptive findings from ˆY (e.g. product demand, near-est customers, eligible suppliers, or inventory levels). For distribution network optimisation, ˆA synthesises an interpretable shipping plan, allocation records, and cost breakdown from ˆY. For what-if analysis, ˆA com-pares baseline and counterfactual scenarios. Let ˆY base and ˆY cf denote results from ˆC base and ˆC cf, respec-tively. The module produces a contrastive explana-tion ˆA = LLLM (Pans(Q, ˆY base, ˆY cf)), highlighting the impact of the counterfactual constraints on decision out-comes (e.g. changed allocations, or cost shifts). This study argues that decision support quality depends on both whether the graph database query is executable to retrieve the correct information within the knowl-edge graph and whether the explanation faithfully reflects the query’s output. Therefore, the correctness and faith-fulness of a pair (ˆC, ˆA) should be evaluated together, preventing two failure modes: a plausible explanation from the user query despite an incorrect ˆC, and a correct ˆC whose results are poorly interpreted. This design motivates a tailored evaluation metric, accurate query-explanation consistency (see Section 4.2), which penalises an imbalance between query correctness and explanation faithfulness. 3.5. Digital twin synchronisation and adaptive. knowledge graph update This module ensures that supply chain digital twins are synchronised with operational changes and the LMiDTSC method’s retrieval memory is adaptively updated to improve future graph database query gener-ations. The subgraph Gt = (Vt, Et, μt, δt) serves as the digital twin artefact for real-time supply chain visibility. In practice, ˆG Q is rendered as a visual graph alongside the final answer ˆA, enabling users to inspect the real-time and structural contexts behind the distribution plans or scenario analyses. The synchronisation aims to main-tain Gt+1 = sync(Gt, 1t), where 1t can be induced by user-initiated actions or changes in the external envi-ronment. For instance, an optimal shipping plan can be exported as a spreadsheet for procurement planning or inventory replenishment. Conversely, when operational changes (e.g. updated inventory records or revised costs) occur as structured files, the LMiDTSC method imports these updates into the supply chain knowledge graph. In addition, the proposed method improves its con-tinuous learning by updating the retrieval memory. When LMiDTSC encounters a previously unseen user request but successfully generates an executable query and a grounded explanation, this resolved case can be inserted as a new exemplar into the vector database to serve future in-context learning prompts. Opera-tionally, a curation rule avoids inserting near-duplicate exemplars into the vector database by checking whether the maximum similarity to existing exemplars stays below a threshold τ: max bi∈B cos(f (Qnew), bi) ≤ τ ⇒ B ← B ∪ (f (Qnew), Qnew, ˆC new). This adaptive update can accumulate domain-specific graph database query exper-tise over time to improve in-context learning effec-tiveness. Besides, it provides persistent documenta-tion by logging the system’s behaviour to create an auditable trail of how the information was queried and what operational decisions were derived. This design advances a digital twin-based decision support system to move beyond solver-centric optimisation-simulation by enabling natural-language interaction and explainable decision support. Unlike conventional digital twin-based decision support systems that require manual reformula-tion of optimisation-simulation models, parameters, and constraints, LMiDTSC maintains a knowledge graph-based digital twin’s state that can be queried and updated through natural-language-driven multi-hop reasoning and synchronisation. The key performance of the pro-posed method compared to the baselines is presented in Figure 2. 4. Empirical setting. This section presents the data collection and processing based on the case study to design a knowledge graph and a multi-task benchmark. It also presents evaluation metrics and implementation details for intensive experi-ments to evaluate the proposed method’s effectiveness in addressing focal CLSC operational tasks. 4.1. Data collection and processing. The case study on the steel sector is used to design the benchmarks and implement the evaluation. The case company operates a country-level forward and reverse distribution network of steel products and scrap in a variety of profiles (e.g. L, U, I, H-shaped sections). This study utilises operational record tables in spreadsheets from the case company to construct a domain-specific graph dataset. This work manually reviews, encodes, and edits the entity name and location, and establishes the relationships between entities. These steps include data cleaning, deduplication, and deidentification to preserve data confidentiality. The graph dataset is composed in Vietnamese and then translated to English to ensure that large language models can accurately understand and capture mean-ing. The dataset comprises 20 suppliers, 5 warehouses, and 3 return centres, owned and operated by the case company, and 500 customers trading 115 steel product types and 6 steel scrap types, along with the relation-ships among these entities (see Tables 1 and 2). This study then constructs a knowledge graph of the steel product and scrap distribution containing 649 entities and 4891 relationships within Neo4j, a market-leading commercial graph database management system (see Figure 3). This approach provides the basis for the benchmarking and evaluation. 4.2. Benchmarks and evaluation metrics. The benchmark focuses on three operational tasks supporting decision-making processes in CLSC operations: supply chain visibility, distribution network optimisa-tion, and what-if analysis. Although the benchmark is grounded in a steel CLSC, it is designed to be trans-ferable to other supply chain settings. Particularly, the optimisation task focuses on forward steel distribution and reverse scrap distribution problems, while the what-if analysis task focuses on analysing counterfactual con-straints around the optimisation settings. Readers can refer to Section 3.3 and Appendix 2 for the model formu-lations, graph-native algorithms, and solution templates. For each task, this work compiles a repository of request-query-explanation triplets (i.e. user request, ground-truth graph database query, and ground-truth explana-tion). This work implements Python macros for base test cases to generate a suite of benchmark instances (i.e. 72 supply chain visibility, 54 distribution network opti-misation, and 34 what-if analysis test cases), known as the BenchSteel-160 benchmark. All test cases are created from scratch and kept in-house. Meanwhile, large lan-guage models used in the proposed method are accessed through application programming interface (API) calls to ensure models cannot observe benchmark instances in advance. To systematically evaluate the proposed method’s per-formance across focal tasks, this study adopts widely used text-generation metrics from natural language process-ing: BLEU (Bilingual Evaluation Understudy), ROUGE (Recall-Oriented Understudy for Gisting Evaluation), BERTScore (Bidirectional Encoder Representations from Transformers Score), and large language model-as-a-judge (LLM-Judge). In addition, this work proposes a tailored metric, termed the accurate query-explanation consistency (AQEC), to jointly assess the correctness of the generated graph database query’s output and the semantic faithfulness of the corresponding explanation. The AQEC is well-suited to evaluate the multi-step rea-soning pipeline in the proposed method and is therefore used as a primary metric throughout empirical findings and discussion. Overall, this study employs BLEU and ROUGE to evaluate the quality of the generated graph database query’s output; BERTScore, Cosine Similarity, and LLM-Judge to evaluate the generated explanation quality; and AQEC as a mixed evaluation metric. The calculation of each evaluation metric is detailed below. AQEC. This study argues that the final explanation should be grounded in the graph database query’s output to minimise the risk of large language models’ halluci-nations. Therefore, the correctness of the query’s output and the explanation quality are treated as equally impor-tant. Because using cosine similarity only may overes-timate explanation quality, as models can repeat salient tokens from the user request while producing an incor-rect query in the previous reasoning step. Inspired by unit testing in software development, AQEC is designed to simultaneously quantify the correctness of a graph database query’s output and the faithfulness of a final explanation. The rationale is that evaluating graph database queries’ outputs is analogous to analysing coding accuracy, but existing evaluation metrics, for example, the correctness accuracy of Li et al. (2023), are insufficient for our setting for three reasons. First, this correctness accuracy metric exhibits weak dis-criminative power across different large language models’ performances. Second, it is designed primarily to eval-uate a simple generation of a few elements in Python code lines, whereas our graph database queries’ outputs are structured dictionaries. Third, it overlooks the expla-nation quality, yet it is critical in our setting because decision-making may be compromised when explana-tions are unclear. Specifically, AQEC integrates fuzzy matching based on Levenshtein distance as defined in Equation and cosine similarity as defined in Equation. AQEC does not require the generated query’s output to exactly match the ground-truth one, as multiple query formulations may yield identical outputs. Let SGen and SGT denote the outputs of generated and ground-truth graph database queries, respectively, and EGen and EGT denote the vector embeddings of generated and ground-truth explanations. Each output dictionary is serialised in canonical form, where keys are sorted alphabetically, and numeric values are normalised to fixed precision prior to fuzzy matching. AQEC is defined in Equation. cos(EGen, EGT) = EGenEGT EGenEGT AQEC = 2 fuzz(SGen, SGT)cos(EGen, EGT) fuzz(SGen, SGT) + cos(EGen, EGT) BLEU. This measures the similarity between the gen-erated text and ground-truth text to evaluate genera-tion quality. Similarity is com-puted via overlap of n-grams between the generated text and the ground-truth text. This study employs BLEU-1 and BLEU-2. Let GGen denote the executed result from the generated graph database query, GGT denote the executed result from the ground-truth graph database query, countGen represent the number of times an n-gram appears in GGen, and countGT represent the number of times the n-gram appears in GGT. For a given n, BLEU-n is computed in Equation. n-grams∈GGen min(countGen, countGT) ∑︁ BLEU-n = ∑︁ n-grams∈GGen countGen ROUGE. It emphasises content coverage, making it particularly suitable for evaluating whether key informa-tion in the executed result of the ground-truth graph database query is preserved in the generated output. This study employs ROUGE - 1 and ROUGE-2, which compute the recall of 1-gram and 2-gram in the executed result of the ground-truth graph database query. For a given n, ROUGE-n is computed in Equation. ∑︁ n-grams∈GGT min(countGen, countGT) ROUGE-n = ∑︁ n-grams∈GGT countGT BERTScore. This measures semantic similarity by matching each token in the generated text with each token in the ground-truth sentence using contextual embeddings. Given a generated token sequence Y = [y1..., yn], and a ground-truth token sequence X = [x1..., xm], the contextual vector embeddings ̃Y = [ ̃y1..., ̃yn], and ̃X = [ ̃x1..., ̃xm] in the explanations are computed by text embedding mod-els. The metric computes recall by matching each vector in ̃X to its most similar vector in ̃Y as defined in Equation, and precision by matching each vector in ̃Y to its most similar vector in ̃X as defined in Equation. Then, BERTScore is the harmonic mean of precision and recall in Equation. RecallBERT = 1 ∑︂ max(̃xmT ̃yn), ∀ ̃yn ∈ ̃Y ̃X ̃xm∈ ̃X PrecisionBERT = 1 ∑︂ max(̃xmT ̃yn), ∀ ̃xm ∈ ̃X ̃Y ̃yn∈ ̃Y BERTScore = 2 PrecisionBERTRecallBERT PrecisionBERT + RecallBERT LLM-Judge. This leverages a large language model to score generated texts against ground-truth texts accord-ing to semantic correctness and helpfulness. It approximates human judgements by captur-ing contextual information while substantially reducing the labour-intensive and time-consuming nature of man-ual judgement. For objectivity and bias reduction, this study employs a Llama-3.3-70B-Instruct model to judge the incorporations of GPT models within the proposed method. The detailed prompt is provided in Appendix 3. 4.3. Implementation details. The experimentation follows best practices in machine learning. This study splits the benchmark into an example set and an evaluation set. The former provides learning examples for in-context learning, and the latter is used to evaluate the out-of-sample performance. This design prevents data leakage and enables systematic evaluation of the generalisability and extrapolation of the proposed method to unseen test cases. This work randomly assigns 90% of the benchmark to the evaluation set and 10% to the example set. Due to the imbalance in the number of tasks, we apply a strati-fied split so both sets preserve the same focal task ratios. To ensure robustness, each experiment is repeated three times and reported with the mean and standard deviation of evaluation metrics on the evaluation set. This study selects three proprietary large language models: GPT-5.2, GPT-4.1, and GPT-3.5, and two open-source models: gpt-oss-20b and Llama-3.1-8B-Instruct for experimentation. For text embedding generation, this study employs a proprietary text-embedding-ada-002 model and an open-source all-MiniLM-L12-v2 model. Llama-3.3-70B-Instruct is employed for LLM-Judge evaluation. The selection is based on the fol-lowing considerations: the availability of API calls, open-source and resource-efficient for further fine-tuning and enterprises’ deployment in consumer hardware, sufficient coding capability for graph database query generation, and an adequate context window to accommodate an incremental increase of learning examples for in-context learning. Table 3 summarises the models and settings. Importantly, this study empirically observed that the performance does not improve noticeably when the number of learning examples exceeds four. Otherwise, additional learning examples can increase computational costs and induce overfitting. Thus, this study decides to evaluate the performance from 0 to 4 learning exam-ples. For fairness and reproducibility, all incorpora-tions of large language models within the proposed method are evaluated under the same experimental pro-tocol. All experiments and fine-tuning are conducted on an Ubuntu 24.04.3 workstation with an Intel Core i5-14600K CPU, an NVIDIA GeForce RTX 5060 Ti GPU, PyTorch 2.9.1, and CUDA 13.0. 5. Empirical findings and discussion. This section reports intensive experiments to evaluate the overall and task-specific performance of the proposed method against baselines, sensitivity to text embedding generators and fine-tuning, and robustness under the semantic diversity of user inputs. Appendix 2 presents detailed examples of ground-truth responses. 5.1. Comparison of the proposed method against. baselines This study now evaluates the performance of the pro-posed method (LMiDTSC) in interpreting, reasoning over, solving and explaining across focal operational tasks (i.e. supply chain visibility, distribution network optimi-sation, and what-if analysis) (see Table 4). The methods are evaluated on their capability to interpret the user request, then generate correct graph database queries’ outputs, and provide an accurate explanation in natural language. To emphasise the practicality of the resource-efficient models, this study uses all-MiniLM-L12-v2 and gpt-oss-20b as the backbone language models within LMiDTSC for this evaluation. This work follows com-mon practices in machine learning and evaluates the performance on out-of-sample observation, that is, the evaluation benchmark has not been part of the in-context learning and is thus unseen to baselines and the proposed method. Overall, this study finds that the proposed method is considerably more accurate than baselines across most evaluation metrics. The baselines (large language mod-els only) represent implicit knowledge and might not be pretrained or fine-tuned with focal operational tasks, which are outperformed by a large margin. The pro-posed method using open-source language models yields an BLEU-1 of 83.68%, BLEU-2 of 83.00%, ROUGE-1 of 86.03%, and ROUGE-2 of 84.33% in generating correct queries’ outputs, Cosine Similarity of 83.09% and LLM-Judge of 7.68 out of 10.00 in providing accurate explana-tions, and AQEC of 82.73% in both capabilities. However, LMiDTSC slightly trails the baseline (GPT-5.2 only) by 2.71 percentage points in BERTScore during the expla-nation phase. This difference may reflect the fact that state-of-the-art models (e.g. GPT-5.2) use broad seman-tic paraphrasing, which is favoured by contextual embed-dings of the BERTScore metric, whereas LMiDTSC is optimised for lexical precision and faithfulness to the ref-erence explanation. Altogether, these results demonstrate the proposed method’s effectiveness, with the incorpo-ration of an open-source text embedding model and an open-weight large language model only in interpreting, Note: Reported is the mean (and standard deviation) performance across three independent experiments (in %), except the LLM-Judge (in 0–10 scale). The best value per metric is highlighted in bold. The absolute improvement per metric of the proposed method over the best baseline is underlined (in %). reasoning over, solving, and explaining focal operational tasks based on the knowledge graph of the steel product and scrap distribution network. 5.2. Performance across focal operational tasks. This study now compares the proposed method’s per-formance when different large language models are used as the backbone across three focal operational tasks (see Table 5). The objective is to evaluate how much LMiDTSC improves when it incorporates state-of-the-art large language models and open-weight mod-els, given that approaches relying solely on implicit knowledge from large language models often fail to address domain-specific tasks. In principle, incorporat-ing high-performing models should improve domain-specific applications because such models have been pretrained on broad knowledge and varied task pat-terns. This study evaluates different large language mod-els within LMiDTSC on three subsets of the out-of-sample evaluation benchmark, grouped by operational task type, and across different numbers of learning exam-ples used for in-context learning. This work employs AQEC to quantify the evaluation. Overall, this work finds that performance is substantially robust. Incorpo-rating state-of-the-art GPT-5.2 and GPT-4.1 models into LMiDTSC yields higher performance than the legacy GPT-3.5 model and the open-weight gpt-oss-20b model across all operational tasks. Specifically, the AQEC increases as the number of in-context learning examples increases across methods. However, the performance does not improve significantly from two or three learning examples. Interestingly, this pattern is also visible for state-of-the-art models. For instance, incorporating the GPT-4.1 model leads to a decrease of 0.54 percentage points in supply chain visi-bility tasks and 4.01 percentage points in what-if analysis tasks when moving from two-shot to four-shot settings. On the other hand, all incorporations exhibit strong performance in supply chain visibility tasks and strug-gle substantially in optimisation and what-if analysis tasks when only zero or one learning examples are pro-vided. These results can be explained by differences in task complexity. Supply chain visibility tasks primar-ily involve straightforward information extraction and aggregation, whereas optimisation and what-if analy-sis tasks require multi-step reasoning and the genera-tion of syntactically correct queries that involve complex mathematical operations over multiple entities and rela-tionships. For example, in supply chain visibility tasks, identifying nearest customers requires the correct use of the Haversine formula for spherical distance calculation, where the zero-shot setting often fails but improves after a few-shot exposure. By contrast, distribution optimisa-tion and what-if analysis involve long query sequences with multiple constraints, which impose higher logical demands and therefore explain the observed underper-formance (see Appendix 2). For example, under zero and one learning example settings in optimisation tasks, most incorporated models tend to simply extract enti-ties and distribution routes mentioned in the user request to directly provide the answer, rather than computing demands, inventory level, and supply sources needed for optimal allocation. In addition, incorporating the GPT-5.2 model yields the strongest performance for most operational tasks under zero-shot in-context learning, reaching 51.85% for supply chain visibility, 26.87% for distribution opti-misation, and 81.01% for what-if analysis. The best performance in each task is observed when incorpo-rating the GPT-4.1 model, wherein 89.32% of sup-ply chain visibility cases are addressed by two learn-ing examples, 91.91% of distribution optimisation cases are addressed by four learning examples, and 93.74% of what-if analysis cases are addressed by two learning examples. Yet incorporating the gpt-oss-20b model still achieves competitive performance, reaching 87.42% for Note: Reported is the mean (and standard deviation) AQEC across three independent experiments (in %). The best value per method, task, and the number of learning examples is highlighted in bold. supply chain visibility, 78.07% for distribution optimisa-tion, and 86.28% for what-if analysis. Using LMiDTSC with an appropriate number of learning examples can further increase the AQEC, which reflects its effective-ness in addressing focal tasks. In particular, with GPT-5.2 or GPT-4.1 models and two learning examples, performance increases by 36.57–49.79% for supply chain vis-ibility tasks, 60.54–63.63% for optimisation tasks, and 8.11–12.81% for what-if analysis tasks, compared to the zero-shot setting. Therefore, our findings suggest that incorporating both proprietary and open-weight large language models within the proposed method is informative for addressing focal tasks and supporting decision-making in steel CLSC operations through a robust generation of graph database queries’ outputs and faithful result explanations. 5.3. Sensitivity to text embedding generator. This study now provides a sensitivity analysis by vary-ing text embedding models within the proposed method, including open-source and proprietary models (see Table 6). The role of an embedding model is to construct a vector database for questions (user requests) in the exem-plar set and to generate text vector embeddings of user requests, which are then compared with questions in the vector database to retrieve top - k most similar learning examples. Overall, this work confirms that the choice of embedding models does not lead to a substantial differ-ence in performance. When replacing an open-source with a proprietary text-embedding-ada-002 embedding model in LMiDTSC (gpt-oss-20b model backbone), this work observes improvements of 2.02 percentage points in BLEU-1, 1.53 percentage points in BLEU-2, 4.30 per-centage points in ROUGE-1, 2.95 percentage points in ROUGE-2, 4.80 percentage points in BERTScore, 4.49 percentage points in Cosine Similarity, 0.3 percentage points in LLM-Judge, and 4.59 percentage points in AQEC. Importantly, AQEC indicates that changing from an open-source to a proprietary model increases the rate of successfully addressing focal tasks from 82.73% to 87.32%. This improvement is explained by the fact that a larger embedding dimensionality of a proprietary model captures richer semantic representation of user requests, therefore, retrieves more relevant learning examples from an exemplar set. Altogether, these results demonstrate the efficacy of LMiDTSC even when using an open-source embedding model and an open-weight large language model in addressing focal tasks. This further supports the managerial implication that implementing open-source language models within the proposed method on con-sumer hardware is feasible without materially compro-mising performance in CLSC operational decision sup-port. This study further examines performance across focal tasks using a varying number of learning examples under different embedding models (see Table 7). Since the role ∗∗Two learning examples. Note: Reported is the mean (and standard deviation) performance across three independent experiments (in %), except the LLM-Judge (in 0–10 scale). The best value per metric and method is highlighted in bold. The absolute improvement of the use of proprietary vs. open-source embedding model per metric is underlined (in %). Note: Reported is the mean (and standard deviation) AQEC across three independent experiments (in %). The best value per number of learning examples, method, and task is highlighted in bold. The absolute improvement of the use of proprietary vs. open-source embedding model per number of learning examples is underlined (in %). of embedding models is to generate vector embeddings of user requests for retrieving top-k similar learning exam-ples, this study does not report a zero-shot setting in this experiment. Importantly, when the number of learning examples increases beyond two, the incorporation of a proprietary embedding model tends to yield better per-formance. One possible explanation is that retrieving a larger set of learning examples places greater demands on semantic matching; the proprietary model may represent user requests more effectively, increasing the likelihood of retrieving more relevant in-context learning exam-ples and enabling LMiDTSC to generate more accurate answers. Indeed, incorporating an open-source model outperforms the proprietary one in settings with fewer learning examples, where 86.84% and 87.42% of visibil-ity tasks are successfully addressed under one-shot and two-shot settings, respectively, and 74.63% of optimisa-tion tasks are successfully addressed under the one-shot setting. However, the opposite pattern is observed in settings with more complex tasks (e.g. optimisation and what-if analysis). Incorporating a proprietary model consistently outperforms the open-source one in the three-shot set-ting, where performance reaches 81.28% AQEC for opti-misation and 88.67% AQEC for what-if analysis. Interest-ingly, for what-if analysis tasks, the performance of using a proprietary embedding model decreases by 3.99% when moving from three-shot to four-shot in-context learning. A similar saturation effect is observed with the incorpo-ration of the open-source embedding model, where per-formance does not improve beyond one learning exam-ple. This may happen because what-if analysis requires the proposed method to edit graph database queries asso-ciated with optimisation tasks to generate counterfac-tual scenarios; thus, additional learning examples do not necessarily provide new useful signals and may instead introduce distracting patterns that reduce generalisation. Therefore, our findings suggest that implementing open-source embedding models within the proposed method is effective and economically attractive for simpler tasks (e.g. supply chain visibility) when fewer learning exam-ples are sufficient. In contrast, proprietary models should be considered for complex tasks (e.g. optimisation and what-if analysis) to retrieve more relevant learning exam-ples for in-context learning to achieve higher accuracy. 5.4. Sensitivity to fine-tuning our proposed method. This study now experiments with fine-tuning the pro-posed method. Specifically, this study fine-tunes the text-generation capability of an Llama-3.1-8B-Instruct model to align with two requirements in the proposed method: graph database query generation and result explanation across focal operational tasks. Dataset preparation fol-lows standard practices in machine learning. This work randomly splits the 90% evaluation set of the BenchSteel-160 benchmark (see Section 4.2) into training and validation sets. The former is used for fine-tuning, and the latter is used to monitor learning progress and to prevent overfitting. This work randomly assigns 90% of samples to train-ing and 10% to validation, and applies a stratified split so both sets contain instances from focal tasks. This work further duplicates and processes data into two subsets to fine-tune two adapters: query generation, where the model learns to produce the syntactically correct query, given a user request; and explanation generation, where the model learns to produce logical and succinct explana-tions in a business and management style. Compared with no fine-tuning, our fine-tuned LMiDTSC directly injects domain-specific task structure into large language models, rather than relying only on their implicit knowl-edge. Therefore, our fine-tuning objective is to improve performance with less reliance on in-context learning, while enabling local deployment on consumer hardware and reducing hallucination risks caused by irrelevant pre-trained knowledge in the context of steel CLSC opera-tions. This study conducts supervised fine-tuning (SFT) for a Llama-3.1-8B-Instruct model within LMiDTSC using two adapters: graph database query generation and result explanation. The SFT learns task-specific behaviour by minimising a next-token prediction loss on labelled query-response pairs. The SFT objective is to minimise the loss computed only over output tokens, as shown in Equation. LSFT [︄ ]︄ ∑︂ log p(xi|x0, x1..., xi−1; 2) = Ex∼DSFT − i∈O Here, 2 denotes trainable parameters, DSFT represents the fine-tuning dataset, O denotes the set of output token positions, x = (x0, x1..., xn) is the tokenised sequence comprising both input prompt and target output, and n is the index of the final token in the sequence. The sum in Equation runs over output token positions only, so the loss is computed on the model’s prediction of each output token xi conditioned on all preceding tokens (x0, x1..., xi−1). However, a full SFT requires substan-tial computational resources and sometimes is unneces-sary as the domain-specific dataset is relatively limited. This work employs the parameter-efficient fine-tuning (PEFT) method. This work adopts QLoRA, which com-bines low-rank adaptation (LoRA) with 4-bit quantisation of the base model, within the PEFT setup. The LoRA technique employs low-rank decompo-sition to parameterise weight updates. For a pretrained weight matrix in the foundational model W0 ∈ Rd×k, the weight update is defined in Equation. W0 + 1W = W0 + (α/r)BA Here, B ∈ Rd×r, A ∈ Rr×k, rank r ≪ min(d, k), and α controls the magnitude of the update. During training, W0 remains frozen, and only A and B are trainable. The modified forward pass for an input vector x is defined in Equation. This approach updates a small set of train-able parameters while keeping the base model weights fixed, which substantially reduces memory usage and training cost, and is suitable for training and inference on consumer hardware. The setup and configuration of SFT are presented in Table 8. h = W0x + 1Wx = W0x + (α/r)BAx Table 9 reports the results. For the no-fine-tuning candidate, this work uses the best-performing base-line setting from earlier experiments, that is, LMiDTSC with GPT-5.2 as the backbone with four learning exam-ples. Overall, this study observes consistent improve-ments across all evaluation metrics after fine-tuning. Specifically, the fine-tuning yields increases of 0.64 per-centage points in BLEU-1, 0.59 percentage points in BLEU-2, 0.41 percentage points in ROUGE-1, 0.34 per-centage points in ROUGE-2, 3.61 percentage points in BERTScore, 5.47 percentage points in Cosine Similarity, 0.63 percentage points in LLM-Judge, and 3.60 percent-age points in AQEC, compared with the baseline. These results indicate that a compact, open-weight large lan-guage model, when adapted to domain-specific tasks, can outperform the use of a non-fine-tuned state-of-the-art model. This supports the view that general-purpose large language models may not be exposed to specialised rea-soning patterns and domain knowledge required by stud-ied tasks in steel CLSC operations, whereas fine-tuning efficiently aligns generation behaviour with the required query and explanation structures. In addition, this study further evaluates the performance of the fine-tuned proposed method for each focal task in more detail (see Table 10). Due to computa-tional constraints during inference, this work reports the performance of fine-tuning under zero-shot and one-shot settings. Even under this restricted evaluation, the findings remain informative. Overall, the fine-tuned proposed method outperforms the non-fine-tuned one across all focal tasks. In a zero-shot setting, the fine-tuned LMiDTSC addresses up to 83.46% of visibility, 50.57% of optimisation, and 92.81% of what-if analysis tasks. Relative to the corresponding non-fine-tuned settings, these results represent improvements of 31.61, 23.70, and 11.80 percentage points, respectively. Interestingly, fine-tuning under a one-shot setting also exceeds the strongest non-fine-tuned results, even those using more learning examples. Specifically, fine-tuning yields a 5.20 and a 1.93 percentage points increase in visibility and optimisation tasks, respectively, compared to the non-fine-tuned base-line using four learning examples. For what-if analysis, fine-tuning yields modest gains over the strongest non-fine-tuned setting (one-shot), with improvements of 1.21 percentage points under zero-shot and 0.07 percentage points under one-shot. The findings underline an impor-tant methodological implication, that is, employing high-performing, general-purpose large language models with intensive in-context learning does not guarantee superior performance on domain-specific tasks, whereas targeted fine-tuning can deliver substantial gains with signifi-cantly less in-context learning efforts. 5.5. Robustness check across the semantic diversity. of inputs This study now evaluates the performance of the pro-posed method when user requests are expressed with higher semantic diversity in natural language. This set-ting is important for practical deployment because supply chain practitioners may not follow a consistent phrasing style when posing requests. This work, therefore, expands the evaluation benchmark by creating paraphrased vari-ants of each test case. Focus groups with linguistic experts and supply chain researchers were conducted to para-phrase questions into a business and management style while preserving their original meanings. For example, a what-if analysis request, such as ‘What would happen if the demand of customer ‘customername’ for the prod-uct ‘productname’ increased by ‘valuenumber’%?’, can be paraphrased as ‘What would be the impact if cus-tomer ‘customername’ increased their demand for prod-uct ‘productname’ by ‘valuenumber’%?’. Each original test case is paraphrased into four additional variants that share the same ground-truth graph database query, query’s output, and output’s explanation. This process yields 360 supply chain visibility test cases, 270 distri-bution network optimisation test cases, and 170 what-if analysis test cases, known as the BenchSteel-800 bench-mark. The performance of the proposed method with different large language models as backbones across the two benchmarks is then reported (see Table 11). In the BenchSteel-160, the proposed method with GPT-5.2 achieves the strongest query output generation performance, reaching 90.87% for BLEU-1, 90.25% for BLEU-2, 92.71% for ROUGE-1, and 91.13% for ROUGE-2 under a four-shot setting. In contrast, LMiDTSC with GPT-4.1 stands out in explanation performance, with 93.63% in BERTScore and 90.52% in Cosine Sim-ilarity under a four-shot setting. When user requests are expressed in more diverse ways (BenchSteel-800), incorporating GPT-4.1 becomes the most robust back-bone across most evaluation metrics, with 90.31% in BLEU-1, 89.61% in BLEU-2, 92.28% in ROUGE-1, 90.63% in ROUGE-2, 93.41% in BERTScore, 90.18% in Cosine Similarity, and 89.86% in AQEC, compared to an incorporation of GPT-5.2. Interestingly, compared with BenchSteel-160, LMiDTSC slightly improves in the query’s output generation performance. For example, the incorporation of GPT-4.1 as the backbone improves 0.82-0.97 percentage points, while the incorporation of GPT-3.5 as backbone improves 0.39-0.93 percentage points, and explanation performance remains broadly stable for ∗One learning example, ∗∗∗∗Four learning examples. Note: The method without fine-tuning uses LMiDTSC with GPT-5.2 as the backbone. Reported is the mean (and standard deviation) performance across five experiments (in %). The best value per metric and method is highlighted in bold. The absolute improvement of the fine-tuning over the no fine-tuning per metric is underlined (in %). Note: The method without fine-tuning uses LMiDTSC with GPT-5.2 as the backbone. Reported is the mean (and standard deviation) AQEC across three independent experiments (in %). The best value per the number of learning examples, method, and task is highlighted in bold. The absolute improvement of the fine-tuning over the no fine-tuning per the number of learning examples is single-underlined (in %). The absolute improvement of the fine-tuning over the best performance of the no fine-tuning is double-underlined (in %). The best performance of the no fine-tuning per task is wave-underlined (in %). ∗∗Two learning examples, ∗∗∗Three learning examples, ∗∗∗∗Four learning examples. Note: Reported is the mean (and standard deviation) performance across three independent experiments (in %). The best value per metric and method in each benchmark is highlighted in bold. The absolute change in the performance between BenchSteel-160 and BenchSteel-800 for the same method is underlined (in %). Note: Reported is the mean (and standard deviation) AQEC across three experiments (in %). The best value per the number of learning examples and method in each benchmark is highlighted in bold. The change in the performance between original and paraphrased benchmark within the same method is underlined (in %). the choice of GPT-4.1. By contrast, the incorporation of GPT-5.2 exhibits a modest reduction on BenchSteel-800 across most metrics, including a decrease of 1.04 percent-age points in AQEC, while standing out with 8.06 out of 10 in LLM-Judge. Overall, the paraphrased benchmark, BenchSteel-800, does not cause significant drops in per-formance, especially with high-performing backbones, which verifies the robustness of the proposed method under semantically diverse user inputs. The study further examines the performance of the proposed method by the number of learning exam-ples across both benchmarks using the AQEC met-ric (see Table 12). Compared to BenchSteel-160, the effect of increasing the number of learning examples on BenchSteel-800 is not uniformly beneficial and some-times shows diminishing returns. For example, in the BenchSteel-800, with the GPT-5.2 backbone, AQEC slightly drops 0.04 percentage points, from 88.53% in the three-shot to 88.49% in the four-shot in-context learning. A similar saturation pattern is visible for other backbones in specific settings, indicating that additional exemplars may introduce redundancy rather than providing use-ful signals for in-context learning. Altogether, this study finds that semantic diversity in user requests has a lim-ited, but non-negligible effect on the overall performance of LMiDTSC. Still, robustness is best maintained by pair-ing LMiDTSC with a high-performing large language model as the backbone and a moderate number of learn-ing examples, typically in two-shot to three-shot set-tings. This experiment verifies the robustness of the pro-posed method across variations in how users phrase their requests on CLSC operational tasks. 6. Implications. The proposed method, LMiDTSC, integrates large lan-guage models and digital twins to support decision-making through generating accurate and interpretable answers for supply chain visibility, distribution network optimisation, and what-if analysis tasks. It enables inter-active analysis, optimisation, visualisation and synchro-nisation given natural-language requests without requir-ing repetitive optimisation-simulation model formula-tion, manual analytics, and result communication. In turn, these enable computerised decisions, strengthening closed-loop operations, stakeholder collaboration, and sustainable performance in the steel sector. Based on our findings, theoretical, methodological, and practical implications are discussed below. 6.1. Theoretical implications. This study contributes to theory on decision support by elucidating how modern decision support systems should process information to translate managerial intent into reliable operational actions. This contribution is grounded in the Organisational Information Processing Theory, which suggests that as task complexity increases (e.g. distribution in CLSCs), organ-isations must enhance their information-processing model in decision-making processes to achieve a given level of performance. While traditional models view information processing through the lens of human intel-ligence (i.e. input, processing, and storage), this study extends that metaphor to modern organisational information systems. This study proposes that an intelligent digital twin-based decision support system can serve as an extended working mem-ory for organisations: real-time, heterogeneous data (sen-sory memory) is embedded into a supply chain knowl-edge graph (long-term memory) to support operational decisions (short-term memory). Furthermore, incorpo-rating large language models shifts the nature of the pro-cessing paradigm from rule-based approaches and rigid algorithms to context-aware interpretation, effectively reducing cognitive load for supply chain practitioners. Consequently, the ‘Intelligence-Design-Choice’ decision-making model, particularly in operations and supply chain management, can be restructured: intel-ligence is automated by large language models that pro-cess unstructured data from the environment, design is supported by digital twin-based course of actions, and choice is guided by multi-step reasoning over the supply chain knowledge graph. From an operations management perspective, a digital twin can be more than a computational and information exchange model. It can interpret natural-language inputs through large language models, reason over a structured enterprise knowledge base to generate outputs, and communicate results through rigorous explanations. Accordingly, the proposed decision support method extends established definitions and architectures of intelligent digital twins by integrat-ing natural-language interpretation, advanced reasoning, and explanation layers. Furthermore, this work advances the human-AI collaboration paradigm, wherein supply chain practitioners interact with large language model-augmented digital twins through a sequence of follow-up queries until they obtain sufficient insights for making decisions. Importantly, these insights are not exclusively derived from general-purpose large language models, but are firmly anchored in enterprise information systems and supply chain knowledge graphs. Supported by digital twins, human-AI collaborative sys-tems extend from data analytics towards transformative actions that reconfigure physical supply chain resources, a progression firmly grounded in the Dynamic Capabil-ities Theory. Finally, this work reinforces operational values of a synergistic AI/GenAI-digital twin system in converging physical and digital organisational resources to drive sustainable transitions in operations and supply chain management. 6.2. Methodological implications. The study contributes to digital twin-based decision support and operations and supply chain management research by demonstrating the value of large language models in enabling more effective CLSC operational decisions. Thereby, it connects to a growing stream of research on incorporating large language models into digital twins, and on large lan-guage models for supply chain optimisation and analytics. Different from prior work that employs solely large language mod-els that retrieve triplets or quintuplets from knowledge graphs to answer relatively simple supply chain visibility questions, the proposed method can reason over a knowledge graph by generating graph database queries that retrieve information from multi-hop entities and relationships to address more complex tasks (e.g. distribution net-work optimisation and what-if analysis). Also, different from prior work that employs large language models to edit linear programming-based optimisation models in a source-code editor to generate counterfactual scenarios for shipping plans, this work adapts lin-ear programming-based optimisation and what-if anal-ysis models into graph database queries, and utilises graph-based retrieval-augmented generation to provide answers evaluated by a tailored evaluation metric. In other words, this work aims to model and solve supply chain optimisation and what-if analysis problems in the form of knowledge graphs and graph database queries. This work speeds up the modelling, reasoning and infer-ence phases based on the effective scale-up nature of knowledge graphs, rather than editing optimisation and what-if analysis code lines as in prior work, while fitting the nature of CLSC optimisation-simulation and analyt-ics problems, where multiple entities and relationships can be modelled by graph representation. Incorporating large language models in digital twin-based decision support systems offers several advantages over single-technology approaches. This integration can interpret natural-language user requests, execute highly adaptive reasoning, and provide tailored insights. On the other hand, this integration opens new opportu-nities to realise a comprehensive concept of digital twins capable of real-time monitoring, virtualisation, simulation-optimisation, and cyber-physical synchroni-sation. In par-ticular, this work provides an approachable method for interactive optimisation, what-if analysis, visualisa-tion, and synchronisation for supply chain monitor-ing, optimisation-simulation, and control operating on both structured data and text. Furthermore, digital twin-based decision support can benefit from fine-tuned large language models by reducing in-context learning effort and operational cost associated with proprietary models. As such, this study expects that the proposed method can provide reliable insights for supply chain operational decision-making with minimised hallucina-tions and without exhaustive in-context learning typical of general-purpose large language model usage. How-ever, applications of large language models in digital twin-based decision support, especially in CLSC oper-ations, are still rare, whereas this work develops a tai-lored method for the studied decision-making problems. Therefore, this study expects that the proposed method is of direct relevance for many supply chain optimisation-simulation and analytics settings where the goal is to integrate digital twins, fine-tuned lightweight, open-source language models, and domain-specific knowledge graphs. This study elucidates the implications of integrat-ing large language models, knowledge graphs, and dig-ital twins within decision support systems. This eval-uation is based on proposing a tailored large language model-augmented intelligent digital twin, a recent inno-vation bridging AI/GenAI and operations and produc-tion research. This work expects that large language mod-els are beneficial for a wide array of supply chain opera-tional and managerial decision-making tasks. This opens new research opportunities for adapting large language model-augmented digital twins and knowledge graphs to address other strategic, tactical, and operational deci-sions in CLSCs (e.g. network design, supplier selection, production planning, and vehicle routing). 6.3. Practical implications. From a managerial perspective, our work demonstrates that the proposed method, LMiDTSC, can address CLSC operational tasks, with particular relevance to steel distri-bution and scrap collection. While prior literature shows that large language models and digital twins can sup-port network optimisation and analytics through simple information extraction and retrieval-augmented gener-ation, our findings show that integrating large language models can further enhance intelligent digital twin-based decision support for CLSC operations through multi-step reasoning over a knowledge graph. Here, this study finds that incorporating open-source embedding and open-weight large language models within LMiDTSC increases the AQEC by 34.41 percentage points, achieving an AQEC of up to 82.73%, compared to the best baseline using implicit knowledge of GPT-5.2. As such, our work is of direct managerial relevance, as it provides comput-erised decision support for supply chain practitioners to make operational decisions based on real-time insights, without computer science expertise, while providing an integrated tool for small and medium-sized enterprises (SMEs) with limited resources when operating CLSCs. The study also shows that incorporating state-of-the-art models (e.g. GPT-5.2 and GPT-4.1) within the proposed method yields even stronger performance, addressing up to 89.32% of visibility, 91.91% of opti-misation, and 93.74% of what-if analysis tasks, com-pared to incorporating legacy and open-weight models. However, while GPT-4.1 is stronger in visibility tasks and GPT-5.2 is stronger in what-if analysis tasks, this study observes a competitive performance in optimisa-tion tasks. Importantly, increasing learning examples for in-context learning does not always increase the performance of either proprietary or open-weight models. This indicates that increasing learning examples does not nec-essarily guarantee performance improvements; in con-trast, it requires higher computational cost and can dilute learning signals within in-context learning. Further-more, incorporating proprietary embedding models does not significantly improve the performance compared to open-source ones. Thus, supply chain practitioners should consider using state-of-the-art large language models as backbones within LMiDTSC when dealing with complex operational decision problems, while using legacy and open-source models for simpler operational decisions. Supply chain practitioners should consider using one or two learning examples for in-context learn-ing within LMiDTSC to achieve robust performance. These approaches help businesses maintain the effective implementation of the proposed method and optimise computational resources. Importantly, fine-tuning an open-source, lightweight large language model within the proposed method sub-stantially increases the performance. Fine-tuning, com-bined with a one-shot setting, achieves an AQEC of up to 93.17%, yielding consistent performance improvements of 3.60 percentage points in AQEC, compared to the best configuration that incorporates a GPT-5.2 backbone with four-shot in-context learning. This is meaningful because the proposed method, when incorporating only a fine-tuned model, becomes a plug-and-play solution for CLSC operational decision support systems, requir-ing no additional prompt engineering from supply chain practitioners. Even with a default one-shot setting, sup-ply chain practitioners can address up to 93.84% of visibility, 92.63% of optimisation, and 91.67% of what-if analysis tasks. This finding adds to prior literature that as more and more open-weight, high-efficiency lan-guage models are released, and fine-tuning frameworks become increasingly accessible, fine-tuning should be considered to improve performance in domain appli-cations. Therefore, supply chain practitioners should consider fine-tuning when deploying large language model-augmented digital twin systems locally for accurate, robust, and low-latency performance, while reducing hallucinations of general-purpose models, enabling the use of consumer hardware in inference for data security, and minimising computa-tional costs associated with third-party language model providers. The performance of the proposed method, when incorporating state-of-the-art large language models as backbones, is robust across the semantic diversity of user requests. The results show significantly improved query output generation, while explanation quality decreases only marginally when incorporating the legacy GPT-4.1 model. When users phrase their requests in varying natural-language styles, LMiDTSC consistently main-tains the ability to address 89.86% of all focal tasks, a result comparable to its performance on the non-paraphrased benchmark. We observe a pattern consistent with task complexity: incorporating GPT-5.2 and GPT-4.1 within LMiDTSC maintains performance with only slight decreases of 2.01-3.16 percentage points in AQEC under one-shot setting and 0.80-1.68 percentage points in AQEC under two-shot setting when users’ requests are paraphrased. Furthermore, when users paraphrase their requests, increasing the number of learning exam-ples is unlikely to improve performance; a trend observed across state-of-the-art, legacy, and open-weight mod-els, where performance decreases beyond two-shot or three-shot settings. Accordingly, supply chain practition-ers should consider using state-of-the-art models within the proposed method with one-shot or two-shot settings to maintain performance when dealing with complex operational tasks and semantic diversity in user requests. From a policy perspective, this study directly sup-ports compliance under several regulatory levers shaping the transition towards green steel. First, it can opera-tionalise national and regional scrap utilisation agen-das by enabling real-time measurement and a decision support tool for scrap collection, recycled content, and reverse logistics performance. This is particularly rele-vant to frameworks and directives that prioritise resource recovery and steel recycling, such as Australia’s Circular Economy Framework, the EU’s Cir-cular Economy Action Plan, scrap utilisation targets embedded in China’s 14th Five-Year Plan, and Vietnam’s Action Plan for Circular Econ-omy Implementation (Government of Vietnam 2025). Second, this work strengthens the auditability of sustain-ability disclosure for policymakers by linking operational decisions to reported outcomes and executed enterprise actions. This supports requirements for credible Scope 3 emissions accounting and recycled-content verification, which is critical as public procurement is moving towards low-carbon or recycled-content standards. For instance, Vietnam’s National Green Growth Strategy (Government of Vietnam 2021) aims to increase the green public pro-curement rate to at least 35% of total public procurement by 2030. Third, this work assists corporates in adapt-ing to export restrictions on unprocessed steel scrap, such as those mandated by Australia’s Recycling and Waste Reduction Act for domestic pro-cessing. Finally, it supports markets facing decarbonisa-tion incentives that penalise carbon-intensive steel while rewarding lower-emission production, such as the EU Carbon Border Adjustment Mechanism. As this work is a data-driven tool, policymakers can accelerate its implementation and impact in two prac-tical ways. First, they can foster industry symbiosis and a shared data ecosystem by developing industrial ontol-ogy and standardised reporting templates for steel and scrap flows, emissions factors, and recycled content. This would facilitate data sharing between the indus-try and regulators, thereby reducing compliance friction for policy instruments (e.g. extended producer respon-sibility or carbon border measures). Second, policymak-ers can invest in human capital by bringing AI/GenAI, knowledge graph, and digital twin competencies into professional training, specifically within the steel sec-tor. Enhancing supply chain practitioners’ capability to design and operate virtual decision-support assistants is essential, particularly for SMEs where specialised exper-tise in data analytics and computer science may be lim-ited. 7. Conclusion. The majority of operational decisions in CLSCs involve heterogeneous data and complex interactions among multiple entities and relationships. Accordingly, the decision-making processes of supply chain practition-ers face significant challenges when addressing requests for operational tasks expressed in natural language. To support practitioners, this study developed a large lan-guage model-augmented intelligent digital twin-based decision support method, LMiDTSC, that incorporates textual data and a knowledge graph to provide decision-support insights for CLSC operations. This work demon-strates that the text-generation capability of large lan-guage models provides significant additional reasoning power. This work further proposes a tailored evaluation metric, AQEC, and constructs a specialised knowledge graph and benchmark for our decision-making prob-lems. Through a real-world case study of a steel product and scrap distribution network, the proposed method helps supply chain practitioners automate operational tasks and communicate insights to stakeholders, thereby enabling computerised CLSC operational decision sup-port. Similar to ours, future work could develop large lan-guage model-augmented digital twins to execute decision computation for other supply chain functions. In such scenarios, our findings suggest that combining large lan-guage models and digital twins can significantly enhance the intelligence of decision support systems and improve the accuracy of CLSC operational decision-making. As with other works, this work has several limita-tions. From a data perspective, the steel CLSC graph dataset focuses on a country-level steel product and scrap distribution network, which may limit its appli-cability to a broader scope. The current graph dataset covers a limited range of supply chain entities, rela-tionships, and attributes, which constrains its applica-bility to more varied decision-making problems. From a modelling perspective, while the studied optimisation and what-if analysis tasks focus on distribution net-work and shipping planning, the current formulation does not explicitly integrate other sustainability dimen-sions (e.g. environmental and social objectives). From a method development and evaluation perspective, the benchmark primarily evaluates graph database queries’ output generation and output’s explanation quality; it does not provide a comprehensive comparison of exe-cution time and system-level performance against opti-misation solver-centric approaches (e.g. across different graph sizes, query complexities, or operational loads). Finally, the current method relies on large language model-driven graph database query generation and does not yet consider conventional graph reasoning operators, which may be more efficient for simpler supply chain visibility queries. Future work should expand this graph dataset to incorporate regional and global steel corporations, enabling more comprehensive analysis and benchmark-ing of the proposed method. The dataset should also be enriched with additional entity and relationship types, along with richer attributes, to support a broader range of decision problems, such as strategic decisions (e.g. facility location), tactical decisions (e.g. production planning), and operational decisions (e.g. vehicle routing). Method-ologically, future work should integrate triple-bottom-line dimensions into the optimisation module to capture economic cost, environmental impact, and social con-siderations. Furthermore, the current benchmark should be extended to evaluate whether the proposed method produces accurate visualisation and synchronisation for enterprise information system updates. Future evalua-tions should move beyond accuracy to include system-atic measurements of execution time and resource con-sumption, directly comparing graph-native optimisation and analysis algorithms against solver-centric pipelines across various problem instances. In addition, future research can develop a hybrid reasoning architecture that combines conventional graph reasoning operators (e.g. path-based and neighbour-based search) for sim-ple tasks (e.g. supply chain visibility) with large language model-guided query generation for complex tasks (e.g. optimisation-simulation and what-if analysis) to improve inference speed and computational efficiency. Finally, since the knowledge graph grows through continual updates and case accumulation within vector databases, future work should investigate strategies for effective knowledge management (e.g. graph schema evolution, summarisation, pruning, and indexing) to maintain rea-soning, retrieval, and the system’s continuous-learning capability efficiently while preserving decision traceabil-ity and accountability. CRediT: Chi An Le: Formal analysis, Methodology, Visualiza-tion, Writing – original draft; Devika Kannan: Conceptualiza-tion, Supervision, Writing – review & editing; Kannan Govin-dan: Conceptualization, Supervision, Writing – review & edit-ing; Truong Thinh Nguyen: Data curation, Methodology, Resources Disclosure statement No potential conflict of interest was reported by the author(s). Notes on contributors Chi An, Le received a B.E. degree from Ho Chi Minh City University of Tech-nology, Vietnam, in 2022, and an M.S. degree from the University of Greenwich, United Kingdom, in 2024. He is currently a Ph.D. candidate at the Centre for Sus-tainable Operations and Resilient Supply Chains (CSORSC), Adelaide University, Australia. His research interests include digital twins, large language models, graph neural networks, closed-loop supply chain, and reverse logistics. He received the Higher Degree by Research (HDR) Student Best Paper Award at the Aus-tralian Maritime Logistics Research Network (AMLRN) 2025 Symposium. Devika Kannan is currently the Deputy Director of the Centre for Sustainable Operations and Resilient Supply Chains (CSORSC), Adelaide University, Australia. She has been recognised as a Highly Cited Researcher in ‘Engineering’ (2019, 2021-2024) by Thomson-Reuters/Clari-vate Analytics. She has published more than 90 international journal articles (with 22,840+ citations and an h-index of 62) in leading journals such as Nature, Euro-pean Journal of Operational Research, Omega, Journal of Envi-ronmental Management, Journal of Cleaner Production, Com-puters & Industrial Engineering, Transportation Research Part Author contributions E: Logistics and Transportation Review, International Journal of Production Economics and International Journal of Produc-tion Research. Many of her papers were selected as the ESI top 1% highly cited papers or 0.1% hot papers and highlighted as the Key Scientific Articles contributing to the excellence in Engineering and Environmental research. She is an Editorial Board Member of several international journals. Her research interests include procurement and supply chain management with a strong focus on green, sustainable, circular economies, decarbonisation and digitalised strategies, and sustainable/cir-cular entrepreneurship. Kannan Govindan is currently the Direc-tor of the Centre for Sustainable Oper-ations and Supply Chain Resilience (CSORSC), Chair, Low Risk Research Ethics Committee, College of Business & Law, Adelaide University, Australia, and Chair Professor with the University of Southern Denmark, Denmark. He has been recognised as a Highly Cited Researcher in ‘Engineer-ing’ for eight years in a row (2018, 2019, 2020, 2021, 2022, 2023, 2024, and 2025) by Thomson-Reuters/Clarivate Ana-lytics. He has published more than 475 international journal articles (with 81,100+ citations and an h-index of 150) in leading journals such as Nature, European Journal of Oper-ational Research, Omega, Decision Sciences, Journal of the Operational Research Society, Journal of Environmental Management, Journal of Cleaner Production, Computers & Indus-trial Engineering, Transportation Research Part E: Logistics and Transportation Review, Transportation Research Part D: Transport and Environment, European Management Journal, International Journal of Production Economics and Interna-tional Journal of Production Research. Many of his papers were selected as the ESI top 1% highly cited papers or 0.1% hot papers and highlighted as Key Scientific Articles contributing to excellence in Engineering and Environmental research. He is an executive editor of the Journal of Cleaner Production, Associate Editor of Transportation Research Part E: Logis-tics and Transportation Review, Senior Editor of Industrial Management and Data Systems, Editor of Annals of Opera-tions Research, an area Editor of INFOR: Information Systems and Operational Research, and was a Guest Editor in jour-nals such as Decision Sciences, European Journal of Operations Research, Computers and OR, Annals of OR, the Journal of Cleaner Production, and the International Journal of Produc-tion Economics. Finally, he is an Editorial Board Member of several international journals. His research interests include digital supply chain, Industry 4.0 on supply chain, Sustainable Development Goals, reverse logistics, closed-loop supply chain, digitalised sustainable circular economy, green supply chain management, and sustainable supply chain management. Truong Thinh, Nguyen is currently a director at the Duc Hung Thinh Steel Trading Co., Ltd., Vietnam. He received his B.E. degree from FPT University – Ho Chi Minh City, Vietnam, in 2022. His research interests include large lan-guage models, computer vision, human pose estimation, and digital signal processing. Data availability statement Due to the confidentiality of the company in the case study, supporting data is not available. Appendices. Appendix 1. System prompts Figure A1. System prompt for the instruction of graph database query generation. Appendix 2. An example of the reasoning steps of the proposed LMiDTSC method During the operational decision-making in steel product and scrap distribution, supply chain practitioners may start by raising a supply chain visibility request as follows: “What is scrap currently ordered by supplier “Maruichi Sun Steel JSC” in Tien Giang province?” • Step 1: As shown in Figure A3, the proposed LMiDTSC method generates a graph database query for the user’s request based on the system prompt. Figure A2. System prompt for the instruction of explanation generation. Figure A3. Generated graph database query for the supply chain visibility task. • Step 2: As shown in Figure A4, the proposed LMiDTSC method generates an explanation for the user’s request based on the follow-up system prompt. • The output result is shown in Figure A5. Next, supply chain practitioners may raise a distribution optimisation request for shipping planning as follows: “Please state the optimal shipping plan that sells scrap “Broken D Steel Bar” to supplier “Maruichi Sun Steel JSC” in Tien Giang province.” • Step 3: As shown in Figure A6, the proposed LMiDTSC method generates a graph database query based on the system prompt. • Step 4: As shown in Figure A7, the proposed LMiDTSC method generates an explanation based on the follow-up system prompt. • The output result is shown in Figure A8. Then, supply chain practitioners may want to understand a counterfactual scenario by raising a what-if analysis request as follows: “What would happen if shipping cost for scrap “Broken D Steel Bar” from customer to return facility increased by 25%, and from return facility to supplier “Maruichi Sun Steel JSC” in Tien Giang province now tripled?” Step 5: As shown in Figure A9, the proposed LMiDTSC method generates a graph database query based on the system prompt. Step 6: As shown in Figure A10, the proposed LMiDTSC method generates an explanation based on the follow-up system prompt. The output result is shown in Figure A11. Finally, supply chain practitioners may request a real-time visualisation for monitoring (e.g. supply chain entities and their relationships related to prior optimisation and what-if analysis), and further cyber-physical synchronisation. • Step 7: As shown in Figure A12, the proposed LMiDTSC method generates a graph database query for information extraction within the supply chain knowledge graph based on the prior graph database queries from optimisation and what-if analysis tasks. • The output result is shown in Figure A13. Figure A4. Result from executing the generated graph database query in Figure A3. Figure A5. Generated explanation of result in Figure A4. 36 C. A. LE ET AL. Figure A8. Generated explanation of result in Figure A7. Figure A9. Generated graph-native greedy what-if analysis query for steel scrap distribution optimisation. Note: The specific change compared to the optimisation query (see Figure A6) is highlighted in green. Figure A11. Generated explanation of result in Figure A10. Figure A12. Generated graph database query for subgraph extraction, enabling digital twin-based visualisation. Note: The specific change compared to the what-if analysis query (see Figure A9) is highlighted in green. Appendix 3. The prompt template for LLM-judge Figure A14. Prompt template for LLM-Judge.