Large Language Model-Enhanced Multicriteria Decision-Making for Supplier Selection: A Validated Framework with Statistical Analysis

ABSTRACT This study presents a validated multicriteria decision-making (MCDM) framework for supplier selection that integrates large language models (LLMs) to enhance the evaluation process. Traditional supplier selection methods often rely on simplified scoring mechanisms that may overlook nuanced supplier characteristics. We propose a hybrid approach combining weighted scoring methods with DistilGPT-2 to generate interpretable supplier assessments. Using a data set of 362 Vietnamese textile and apparel companies with 64 evaluation features, our method calculates composite scores based on six key performance categories: cost, delivery, quality assurance, service, customer service, and sustainability. The LLM component processes these numerical scores to generate contextual analyses that aid decision-makers in understanding supplier strengths and weaknesses. Validation against expert human evaluations shows 78.3 % agreement in selection decisions, with Cohen’s κ = 0.672 indicating substantial inter-rater reliability. Statistical analysis reveals significant correlations between sustainability and quality assurance (r = 0.543, p < 0.001), and delivery performance with overall supplier acceptability (r = 0.621, p < 0.001). A comparison with the traditional technique for order preference by similarity to the ideal solution and analytic hierarchy process methods demonstrates superior interpretability while maintaining comparable decision accuracy. Results show that 23 suppliers (6.4 %) met selection criteria, whereas 339 (93.6 %) required further evaluation, highlighting the complexity of supplier assessment. This work demonstrates how LLMs can augment traditional MCDM approaches by providing interpretable narratives alongside quantitative metrics.