Summary of Better Instruction-following Through Minimum Bayes Risk, by Ian Wu et al.
Better Instruction-Following Through Minimum Bayes Risk
by Ian Wu, Patrick Fernandes, Amanda Bertsch, Seungone Kim, Sina Pakazad, Graham Neubig
First submitted to arxiv on: 3 Oct 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 The proposed Minimum Bayes Risk (MBR) decoding method leverages large language model (LLM) judges for supervising instruction-following LLMs, improving test-time performance. By using reference-based LLM judges to select high-quality outputs, MBR decoding outperforms greedy and best-of-N decoding methods on AlpacaEval and MT-Bench datasets. The approach demonstrates consistent gains across LLMs with up to 70 billion parameters, showcasing the potential for smaller LLM judges to supervise larger models. Furthermore, iterative self-training using Direct Preference Optimisation leads to significant performance improvements, matching or exceeding the performance of base models with MBR decoding. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This research paper explores a new way to make large language models better at following instructions. Instead of just picking the first answer that comes to mind, the model uses another AI to help it choose the best response. This approach is called Minimum Bayes Risk (MBR) decoding and it makes the model much more accurate. The researchers tested MBR decoding on two different datasets and found that it worked really well. They also experimented with teaching the model how to get even better, by having it learn from its own mistakes. The results show that this approach can make the model even more accurate than usual. |
Keywords
» Artificial intelligence » Large language model » Self training
