Summary of Fine-tuning Language Models with Reward Learning on Policy, by Hao Lang et al.
Fine-Tuning Language Models with Reward Learning on Policy
by Hao Lang, Fei Huang, Yongbin Li
First submitted to arxiv on: 28 Mar 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
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 Reinforcement Learning from Human Feedback (RLHF) approach aligns large language models (LLMs) with human preferences by collecting human preferences, learning rewards, and optimizing policies. While RLHF is effective, fixed reward models may struggle with off-distribution data when policy optimization shifts the LLM’s data distribution. To address this, the authors propose Reward Learning on Policy (RLP), an unsupervised framework that refines reward models using policy samples to keep them on-distribution. The RLP method combines an unsupervised multi-view learning approach for robust representation learning and a synthetic preference generation approach to simulate high-quality preference data with policy outputs. Experimental results on three benchmark datasets demonstrate the effectiveness of RLP, outperforming state-of-the-art methods. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper is about using computer models that can learn from human feedback. This helps us get computers that are more like humans and can understand what we want them to do. The current way of doing this might not work well when the model changes over time. To fix this, the researchers came up with a new approach called Reward Learning on Policy (RLP). RLP makes sure the computer rewards stay the same even when the model changes. They also developed ways to get better data for training these models. In their tests, they found that RLP worked much better than other methods. |
Keywords
* Artificial intelligence * Optimization * Reinforcement learning from human feedback * Representation learning * Rlhf * Unsupervised
