Online personalizing white-box llms generation with neural bandits

Zekai Chen, Po-Yu Chen, Francois Buet-Golfouse

Abstract: Personalized content generation by Large Language Models (LLMs) in finance presents a challenge: efficiently adapting text to individual preferences without creating unique models for each user. This study introduces an innovative online method for financial applications, employing neural bandit algorithms to dynamically optimize soft instruction embeddings based on user feedback, enhancing personalization in white-box LLMs. Through experiments on public generation tasks, we demonstrate significant performance improvements. Notably, our NeuralTS implementation achieves up to a 62.9% improvement in ROUGE scores and a 2.76% increase in LLM-agent evaluation for personalized content generation. This research showcases the efficacy of neural bandits in refining LLM outputs to align with client-specific needs and regulatory requirements, marking a pivotal step towards feasible and effective adaptive text generation in finance. Our method offers a promising and scalable solution for financial institutions to enhance client engagement, improve risk assessment, and streamline regulatory reporting.