Summary of Rs-dpo: a Hybrid Rejection Sampling and Direct Preference Optimization Method For Alignment Of Large Language Models, by Saeed Khaki et al.
RS-DPO: A Hybrid Rejection Sampling and Direct Preference Optimization Method for Alignment of Large Language Models
by Saeed Khaki, JinJin Li, Lan Ma, Liu Yang, Prathap Ramachandra
First submitted to arxiv on: 15 Feb 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI); Computer Vision and Pattern Recognition (cs.CV); 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 This paper proposes a novel approach to reinforcement learning from human feedback (RLHF) for aligning large language models with user intent. The method combines rejection sampling (RS) and direct preference optimization (DPO) to address the challenges of instability and computational expense in proximal policy optimization (PPO)-based RLHF. The proposed RS-DPO method initiates by developing a supervised fine-tuned policy model, which is then used to sample k responses per prompt. The approach identifies pairs of contrastive samples based on their reward distribution and applies DPO with these samples to align the model with human preference. Experimental results show that RS-DPO effectively fine-tunes language models in limited-resource environments, leading to improved alignment with user intent and outperforming existing methods. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper makes it easier for large language models to understand what people want them to do. Right now, these models can get mixed signals from humans, which makes it hard to make sure they’re doing the right thing. The new method combines two techniques: one that helps the model decide which responses are good or bad, and another that adjusts the model’s behavior based on how well it does. This approach is better than other methods because it can work with limited resources and still gets results that match what people want. |
Keywords
* Artificial intelligence * Alignment * Optimization * Prompt * Reinforcement learning from human feedback * Rlhf * Supervised
