Summary of Iterative Data Smoothing: Mitigating Reward Overfitting and Overoptimization in Rlhf, by Banghua Zhu et al.
Iterative Data Smoothing: Mitigating Reward Overfitting and Overoptimization in RLHF
by Banghua Zhu, Michael I. Jordan, Jiantao Jiao
First submitted to arxiv on: 29 Jan 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (stat.ML)
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 Medium Difficulty Summary: Reinforcement Learning from Human Feedback (RLHF) is a crucial technique for aligning language models with human-centric values. Initially, RLHF learns human values using a reward model from ranking data, but it’s observed that the performance degrades after one epoch and optimizing too much against the learned reward hinders the true objective. This paper addresses these issues by designing an improved reward learning algorithm, ‘Iterative Data Smoothing’ (IDS). IDS updates both the model and the data using soft labels, replacing hard labels. Our findings demonstrate the superior performance of this approach over traditional methods. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Low Difficulty Summary: Imagine a way to make language models more human-like by teaching them what’s good or bad. This paper talks about a problem in making these models learn from people’s feedback. They found that when they try to teach the model, it gets worse after some time. To fix this, they came up with a new way of learning called ‘Iterative Data Smoothing’. It works by updating both the model and the data using soft labels instead of hard ones. The results show that this approach does better than other methods. |
Keywords
* Artificial intelligence * Reinforcement learning from human feedback * Rlhf
