Summary of Enhancing Safety in Reinforcement Learning with Human Feedback Via Rectified Policy Optimization, by Xiyue Peng et al.
Enhancing Safety in Reinforcement Learning with Human Feedback via Rectified Policy Optimization
by Xiyue Peng, Hengquan Guo, Jiawei Zhang, Dongqing Zou, Ziyu Shao, Honghao Wei, Xin Liu
First submitted to arxiv on: 25 Oct 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI); Computers and Society (cs.CY)
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 Rectified Policy Optimization (RePO) to address the issue of “safety compensation” in aligning large language models (LLMs). The current approaches decouple helpfulness and safety, training separate preference models for each. However, this can lead to overly restrictive or unsafe responses. RePO replaces expected safety constraints with critical ones imposed on every prompt, enhancing safety across nearly all prompts. Experiments show that RePO outperforms baseline methods in LLM safety alignment. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Large language models need to balance being helpful and safe. Right now, we train separate models for these two goals, but this can lead to problems. Some responses might be too strict or still not safe enough. To fix this, the authors suggest a new approach called Rectified Policy Optimization (RePO). Instead of setting safety constraints just on average, RePO makes sure every prompt is safe. This helps keep language models safe and helpful. |
Keywords
» Artificial intelligence » Alignment » Optimization » Prompt
