Summary of Exploring and Addressing Reward Confusion in Offline Preference Learning, by Xin Chen et al.
Exploring and Addressing Reward Confusion in Offline Preference Learning
by Xin Chen, Sam Toyer, Florian Shkurti
First submitted to arxiv on: 22 Jul 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI)
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 This paper investigates the limitations of Reinforcement Learning from Human Feedback (RLHF) when trained on data containing spurious correlations. The authors show that offline RLHF is prone to reward confusion, which can lead to unwanted behaviors. To mitigate this issue, they introduce a method that leverages transitivity of preferences and builds a global preference chain using active learning. This approach can significantly reduce reward confusion and improve the overall performance of RLHF models. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper looks at how Reinforcement Learning from Human Feedback (RLHF) works when trained on old data. Sometimes, this data has patterns that don’t actually matter, but the RLHF model might still use them to make mistakes. The researchers show that this can happen a lot and create ways to fix it by using what people prefer and actively learning new things. |
Keywords
* Artificial intelligence * Active learning * Reinforcement learning from human feedback * Rlhf
