Reproduce “Improving LLM-based Recommendation with Self-Hard Negatives from Intermediate Layers” 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
Bingqian Li, Bowen Zheng, Xiaolei Wang, Long Zhang, Jinpeng Wang, Sheng Chen, Wayne Xin Zhao, Ji-rong Wen
ILRec, a framework developed by researchers from Renmin University of China and Meituan, enhances LLM-based recommender systems by dynamically generating token-level 'self-hard negatives' from intermediate layers, leading to improved recommendation quality and training efficiency.
Key insights
- Intermediate layers of an LLM can be leveraged as a dynamic source of fine-grained, self-hard negative signals, which are challenging for the current model to distinguish.
- Both penalizing negative predictions (CPO) and distilling final layer knowledge back to intermediate layers (CPD) are crucial for effective and informative preference learning within LLMs.
- Incorporating external collaborative filtering rewards provides a robust mechanism to regularize against false negatives and inject valuable domain-specific information.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.