Reproduce “Better Rankers, Worse Retrievers: The Hidden Cost of RL Post-Training in Generative Recommendation” 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
Yuheng Chen, Fan Zhang, Yixuan Xu, Chao Yang, Yang Xiang, Can Yang
Generative recommenders can improve top-ten ranking while losing logged targets from deeper candidate lists after reinforcement-learning post-training. On five Amazon-review benchmarks, GRPO raises Office NDCG@10 from 11.59% to 12.55%, but lowers hit@500 from 52.84% to 47.69%. An out-of-rollout anchor raises it to 52.14%, without statistically significant change in top-ten quality; this is benchmark evidence, not an online test.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.