MiniOneRec: An Open-Source Framework for Scaling Generative Recommendation
2510.24431v1
cs.IR, cs.AI
2025-10-30
Авторы:
Xiaoyu Kong, Leheng Sheng, Junfei Tan, Yuxin Chen, Jiancan Wu, An Zhang, Xiang Wang, Xiangnan He
Abstract
The recent success of large language models (LLMs) has renewed interest in
whether recommender systems can achieve similar scaling benefits. Conventional
recommenders, dominated by massive embedding tables, tend to plateau as
embedding dimensions grow. In contrast, the emerging generative paradigm
replaces embeddings with compact Semantic ID (SID) sequences produced by
autoregressive Transformers. Yet most industrial deployments remain
proprietary, leaving two fundamental questions open: (1) Do the expected
scaling laws hold on public benchmarks? (2) What is the minimal post-training
recipe that enables competitive performance?
We present MiniOneRec, to the best of our knowledge, the first fully
open-source generative recommendation framework, which provides an end-to-end
workflow spanning SID construction, supervised fine-tuning, and
recommendation-oriented reinforcement learning. We generate SIDs via a Residual
Quantized VAE and post-train Qwen backbones ranging from 0.5B to 7B parameters
on the Amazon Review dataset. Our experiments reveal a consistent downward
trend in both training and evaluation losses with increasing model size,
validating the parameter efficiency of the generative approach. To further
enhance performance, we propose a lightweight yet effective post-training
pipeline that (1) enforces full-process SID alignment and (2) applies
reinforcement learning with constrained decoding and hybrid rewards. Together,
these techniques yield significant improvements in both ranking accuracy and
candidate diversity.
Ссылки и действия
Дополнительные ресурсы: