Summary of Step-dpo: Step-wise Preference Optimization For Long-chain Reasoning Of Llms, by Xin Lai et al.
Step-DPO: Step-wise Preference Optimization for Long-chain Reasoning of LLMs
by Xin Lai, Zhuotao Tian, Yukang Chen, Senqiao Yang, Xiangru Peng, Jiaya Jia
First submitted to arxiv on: 26 Jun 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 A novel approach to enhancing the robustness and factuality of Large Language Models (LLMs) is proposed, focusing on mathematical reasoning. The challenge lies in ensuring the correctness of each step in a chain of reasoning, which Direct Preference Optimization (DPO) has shown to be limited in addressing. To overcome this limitation, Step-DPO is introduced, treating individual reasoning steps as units for preference optimization rather than evaluating answers holistically. A data construction pipeline is also developed, resulting in a high-quality dataset containing 10K step-wise preference pairs. The findings demonstrate that with minimal training and data, a nearly 3% gain in accuracy can be achieved on MATH for models with over 70B parameters. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary A team of researchers aimed to improve the performance of Large Language Models (LLMs) when doing math problems. They found that traditional methods didn’t work well because they didn’t focus on individual steps in a problem, but instead looked at the whole answer. To solve this issue, they created a new method called Step-DPO that looks at each step in a problem separately. They also built a special dataset to help train these models. The results showed that with very little training and data, these models could get math problems right much better than before. |
Keywords
* Artificial intelligence * Optimization
