Reproduce “Generative Explore-Exploit: Training-free Optimization of Generative Recommender Systems using LLM Optimizers” on OpenResearch

About this paper

Lütfi Kerem Senel, Besnik Fetahu, Davis Yoshida, Zhiyu Chen, Giuseppe Castellucci, Nikhita Vedula, Jason Choi, Shervin Malmasi

A training-free optimization method for generative recommender systems utilizes Large Language Models (LLMs) as optimizers to iteratively improve generated content based on implicit user feedback. It demonstrates that a Generative Explore-Exploit strategy, applied to question generation, significantly increases user engagement (Click-Through Rate) in both e-commerce and general knowledge domains through an LLM-based user simulator.

Key insights

• Large Language Models can effectively function as training-free optimizers for generative recommender systems by interpreting performance feedback embedded directly within their input prompts.
• An Explore-Exploit strategy is crucial for generative recommenders to overcome greedy exploitation, enabling the discovery of latent user preferences and leading to superior long-term user engagement.
• An LLM-based user simulator can reliably model user behavior and generate meaningful implicit feedback (CTR), facilitating rapid and cost-effective development and evaluation of adaptive generative recommendation systems.

Summary by alphaXiv.