Reproduce “ProfiLLM: Utility-Aligned Agentic User Profiling for Industrial Ride-Hailing Dispatch” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tengfei Lyu, Zirui Yuan, Xu Liu, Kai Wan, Zihao Lu, Li Ma, Hao Liu
ProfiLLM introduces an agentic LLM data pipeline for generating utility-aligned user profiles from behavioral logs in industrial ride-hailing dispatch systems. This framework, deployed on DiDi's platform, achieved a +0.47% increase in GMV and reduced various cancellation rates by up to -0.82% in online A/B tests, while maintaining sub-millisecond online latency.
Key insights
- Aligning LLM-generated profiles with downstream prediction utility, rather than merely descriptive fluency, is crucial for achieving performance gains in industrial applications, as naive LLM use can degrade performance.
- A strict offline-online contract is vital for deploying LLM-based systems in latency-sensitive, high-throughput environments like ride-hailing dispatch, by confining LLM inference to offline batch processes.
- Agentic LLMs, augmented with a comprehensive toolkit, can effectively mine global operational knowledge from massive historical data that exceeds typical context windows.
- Adaptive user clustering is essential for addressing long-tail data distributions, allowing robust profile generation for users with sparse interaction histories by aggregating data at the cluster level.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.