Summary of On the Algorithmic Bias Of Aligning Large Language Models with Rlhf: Preference Collapse and Matching Regularization, by Jiancong Xiao et al.
On the Algorithmic Bias of Aligning Large Language Models with RLHF: Preference Collapse and Matching Regularization
by Jiancong Xiao, Ziniu Li, Xingyu Xie, Emily Getzen, Cong Fang, Qi Long, Weijie J. Su
First submitted to arxiv on: 26 May 2024
Categories
- Main: Machine Learning (stat.ML)
- Secondary: Machine Learning (cs.LG); Methodology (stat.ME)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary This paper introduces Preference Matching (PM) Reinforcement Learning from Human Feedback (RLHF), a novel approach that addresses the algorithmic bias inherent in traditional RLHF methods for aligning Large Language Models (LLMs) with human preferences. The proposed method, which leverages the Bradley-Terry-Luce/Plackett-Luce model, incorporates a PM regularizer that balances response diversification and reward maximization. By solving an ordinary differential equation, the regularizer is derived to ensure preference matching. The conditional variant of PM RLHF is tailored for natural language generation and empirically validated on OPT-1.3B and Llama-2-7B models, demonstrating a significant improvement in alignment with human preferences. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper helps make sure big language models work better with what people like. Right now, these models are trained using a way called reinforcement learning from human feedback (RLHF). But this method has a problem: it can lead to some people’s preferences being ignored. To fix this, the authors introduce a new approach called preference matching (PM) RLHF. This method uses a special formula that makes sure the model is fair and takes into account what different people like. |
Keywords
» Artificial intelligence » Alignment » Llama » Reinforcement learning from human feedback » Rlhf
