Summary of Just Say What You Want: Only-prompting Self-rewarding Online Preference Optimization, by Ruijie Xu et al.
Just Say What You Want: Only-prompting Self-rewarding Online Preference Optimization
by Ruijie Xu, Zhihan Liu, Yongfei Liu, Shipeng Yan, Zhaoran Wang, Zhi Zhang, Xuming He
First submitted to arxiv on: 26 Sep 2024
Categories
- Main: Artificial Intelligence (cs.AI)
- Secondary: None
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 proposed online Reinforcement Learning from Human Feedback (RLHF) algorithm focuses on self-rewarding alignment methods for smaller models. To address limitations in current approaches, a novel, only-prompting self-rewarding method is introduced that generates preference datasets without relying on discriminator judgment capabilities. This approach employs fine-grained arithmetic control to optimize the gap between positive and negative examples, generating more hard negatives during training. The algorithm is tested on two base models, Mistral-7B and Mistral-Instruct-7B, achieving 34.5% in Length-controlled Win Rates of AlpacaEval 2.0. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Online RLHF tries to learn from human feedback while using smaller models. A new way to make these models better is proposed by generating datasets based on what humans like and dislike. This method doesn’t rely on expert judgment, making it useful for smaller models. The approach also fine-tunes the algorithm to create harder examples during training. Results show that this method can significantly improve performance. |
Keywords
» Artificial intelligence » Alignment » Prompting » Reinforcement learning from human feedback » Rlhf
