Chenyu Huang, Jianghao Lin, Zhengyang Tang, Bo Jiang, Ruoqing Jiang, Benyou Wang, Lai Wei
A framework combining LLMs and evolutionary search to generate box-type inventory policies with statistical safety guarantees.
Recent LLM-based evolutionary search (e.g., AlphaEvolve) excels on static, structured problems but is not directly suited for dynamic inventory settings requiring online updates. A key challenge is providing theoretical guarantees for the safety and performance of generated policies.
The framework uses a reinforcement learning-trained LLM to process demand data and additional features, generating white-box inventory policies. It introduces confidence-interval-based certification to ensure statistical safety and establishes a unified framework with theoretical guarantees connecting training, inference, and deployment. This allows deriving a lower bound on the probability of evolving a statistically safe and improved policy and characterizing the multi-period performance gap relative to an oracle-safe benchmark.
Tested on synthetic and real-world retail data, InvEvolve outperforms classical inventory policies and deep-learning-based methods. It generates new policies that surpass existing benchmarks in canonical settings, providing both theoretical and practical foundations for LLM-based policy generation in dynamic environments.