AIGP LLM Framework Optimizes E-Commerce Pricing for Long-Term Value

This paper introduces AIGP, a novel framework designed to enhance dynamic pricing in large-scale e-commerce by leveraging Large Language Models (LLMs). Traditional pricing models often lack interpretability, struggle to incorporate unstructured information, and frequently fail to align with crucial long-term business objectives such as cumulative Gross Merchandise Value (GMV), Return on Investment (ROI), and milestone achievement. AIGP aims to overcome these shortcomings. The core of AIGP involves prompting an LLM with domain knowledge, structured data, and textual context to generate interpretable and knowledge-aware pricing decisions. To ensure efficient deployment while maintaining high-quality outputs, the framework employs supervised fine-tuning for knowledge distillation. A critical component is the Long-Term Value Estimator (LTVE), which is trained using offline reinforcement learning on historical data. This LTVE functions as a reward model, scoring candidate pricing actions and generating preference pairs for Direct Preference Optimization (DPO), thereby aligning the pricing policy directly with long-term strategic goals. Extensive evaluations, including offline tests and large-scale online A/B tests conducted on Tao Factory, demonstrate AIGP's significant impact. The framework achieved substantial improvements over the production baseline, reporting a +13.21% increase in GMV, a +7.59% rise in ROI, and an +8.20% improvement in milestone achievement rate over a 14-day period. Crucially, AIGP also provides transparent and interpretable rationales for its pricing decisions, a key advantage over black-box traditional models.