Reproduce “Generative Reasoning Recommendation via LLMs” 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
Minjie Hong, Zetong Zhou, Zirun Guo, Ziang Zhang, Ruofan Hu, Weinan Gan, Jieming Zhu, Zhou Zhao
GREAM, an end-to-end framework, integrates large language models into recommender systems by aligning linguistic semantics with collaborative signals and activating explicit causal reasoning. It introduces high-fidelity item indexing, synthetic reasoning data generation, and a sparse-regularized group policy optimization for stable training, demonstrating improved accuracy in direct recommendation and enhanced interpretability through sequential reasoning.
Key insights
- Aligning LLM's linguistic semantics with collaborative filtering signals, achieved through high-fidelity item indexing and diverse supervised tasks, effectively bridges the semantic gap.
- Activating explicit, multi-stage Chain-of-Thought reasoning through synthetic data and curriculum learning not only enhances interpretability but also positively transfers to direct recommendation performance.
- Sparse-Regularized Group Policy Optimization, with its verifiable residual-sensitive rewards and bonus-calibrated advantage, effectively addresses sparse and stochastic feedback in recommender systems, stabilizing LLM training.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.