Summary of Regularized Best-of-n Sampling with Minimum Bayes Risk Objective For Language Model Alignment, by Yuu Jinnai et al.
Regularized Best-of-N Sampling with Minimum Bayes Risk Objective for Language Model Alignment
by Yuu Jinnai, Tetsuro Morimura, Kaito Ariu, Kenshi Abe
First submitted to arxiv on: 1 Apr 2024
Categories
- Main: Computation and Language (cs.CL)
- 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 Best-of-N (BoN) sampling with a reward model has been demonstrated to be an effective strategy for aligning Large Language Models (LLMs) to human preferences during decoding. However, BoN sampling is susceptible to “reward hacking” when the accuracy of the reward model is not high enough due to the quality or quantity of the preference dataset. To mitigate this issue, we propose MBR-BoN, a variant of BoN that incorporates the Minimum Bayes Risk (MBR) objective as a proximity regularization term at inference time. We empirically and analytically demonstrate that the MBR objective quantifies the proximity of the response to the reference policy, serving as a proximity regularizer. Our experimental results on AlpacaFarm and Anthropic’s hh-rlhf datasets show that MBR-BoN outperforms both BoN sampling and MBR decoding. Additionally, we evaluate MBR-BoN for generating a pairwise preference learning dataset for Direct Preference Optimization (DPO). Empirical results indicate that models trained on a dataset generated with MBR-BoN outperform those with vanilla BoN. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper is about making language models better at understanding what humans like or dislike. Right now, there’s a problem called “reward hacking” that makes the models not very good at this. To fix this, the researchers created a new way to make the models work together with humans. They tested their new method on two big datasets and showed that it worked really well. The results are important because they can help us create more accurate language models that understand what we like and dislike. |
Keywords
» Artificial intelligence » Inference » Optimization » Regularization » Rlhf
