Summary of Towards Improved Preference Optimization Pipeline: From Data Generation to Budget-controlled Regularization, by Zhuotong Chen et al.
Towards Improved Preference Optimization Pipeline: from Data Generation to Budget-Controlled Regularization
by Zhuotong Chen, Fang Liu, Jennifer Zhu, Wanyu Du, Yanjun Qi
First submitted to arxiv on: 7 Nov 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
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 The paper proposes improvements to the Direct Preference Optimization (DPO) pipeline for aligning large language models (LLMs) with human preferences or specific goals. Specifically, it addresses issues with preference data generation and training regularization techniques. The authors demonstrate that existing scoring-based reward models produce poor preference data and perform poorly on out-of-distribution tasks, impacting LLM alignment performance. They propose an iterative pairwise ranking mechanism for generating high-quality preference data and a budget-controlled regularization formulation to improve convergence during training. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper is about making language models work better with what humans want them to do. Right now, it’s hard to get these models to follow human preferences because the way we train them is flawed. The authors figured out that the current method of using rewards from scoring systems doesn’t produce good results and can even make things worse when trying to adapt to new situations. To fix this, they came up with a new way to generate preference data and a special training technique that helps the model learn better. |
Keywords
* Artificial intelligence * Alignment * Optimization * Regularization
