Summary of On Scalable Oversight with Weak Llms Judging Strong Llms, by Zachary Kenton et al.
On scalable oversight with weak LLMs judging strong LLMs
by Zachary Kenton, Noah Y. Siegel, János Kramár, Jonah Brown-Cohen, Samuel Albanie, Jannis Bulian, Rishabh Agarwal, David Lindner, Yunhao Tang, Noah D. Goodman, Rohin Shah
First submitted to arxiv on: 5 Jul 2024
Categories
- Main: Machine Learning (cs.LG)
- 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 In this paper, researchers investigate three protocols for humans to supervise AI: debate, consultancy, and direct question-answering. They use large language models as both AI agents and judges, benchmarking on various tasks with different types of asymmetry. The results show that debate outperforms consultancy in most cases, but not always. Debaters’ ability to choose which answer to argue for increases judge accuracy, while stronger debater models lead to more accurate judgments. This study contributes to the development of scalable oversight protocols for superhuman AI. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper looks at ways humans can supervise really smart computers (AI). They test three ways: one where two AI’s have a discussion, another where one AI tries to convince someone, and finally, one where the person just answers questions. The computer models are used as both the AI’s and the person asking the questions. The results show that when AI’s discuss things, they do better than when one AI tries to convince someone. But, it depends on what kind of task is being done. This research helps us understand how humans can supervise super smart computers. |
Keywords
* Artificial intelligence * Question answering
