ProfiLLM Enhances Ride-Hailing Dispatch with LLM User Profiling

Integrating Large Language Models (LLMs) into industrial ride-hailing dispatch systems to extract semantic features from vast behavioral logs presents a compelling challenge. While traditional matching pipelines rely on structured numerical data, crucial behavioral signals, such as a driver's regional preferences, are inherently contextual and can be naturally captured by LLM-generated user profiles. However, scaling such profiling for live, millisecond-latency dispatchers on platforms with millions of daily orders faces several constraints: log data far exceeds LLM context windows, many users have insufficient interaction history for individual profiling, and fluent profiles don't always translate to improved prediction utility. ProfiLLM addresses these issues with a novel agentic LLM data pipeline. ProfiLLM consists of two main modules: Tool-Augmented Global Knowledge Mining, which equips an LLM agent with analytical tools to extract platform-wide insights and adaptive clustering rules; and Utility-Aligned Profile Exploration, which generates, evaluates, and refines candidate profiles for each cluster using a lightweight utility proxy. Deployed on DiDi's production dispatcher, ProfiLLM achieved substantial improvements in outcome prediction, gross merchandise value (GMV) gains in simulation, and positive results in online A/B tests, including increased completion rates and reduced cancellation rates.