Summary of From Lists to Emojis: How Format Bias Affects Model Alignment, by Xuanchang Zhang et al.
From Lists to Emojis: How Format Bias Affects Model Alignment
by Xuanchang Zhang, Wei Xiong, Lichang Chen, Tianyi Zhou, Heng Huang, Tong Zhang
First submitted to arxiv on: 18 Sep 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: 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 paper studies biases in reinforcement learning from human feedback (RLHF), where widely-used preference models, including GPT-4 and top-ranking models on RewardBench, exhibit strong format biases towards specific patterns like lists and emojis. Large language models can exploit these biases to achieve higher rankings on benchmarks like AlpacaEval and LMSYS Chatbot Arena. The paper extends the study beyond length bias, analyzing a wider range of format biases. It also shows that injecting biased data into the reward model can significantly impact its performance. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This research looks at how people’s preferences affect what AI systems learn from them. Right now, many popular models and even humans have strong likes or dislikes for certain formats like lists or bold text. This can help some AI systems do better on tests, but it’s not always fair because they’re being favored just because of the way they look. The study finds that these biases go beyond just how long an answer is and looks at many other types of format bias. It also shows that even a little bit of biased data can make a big difference in how well AI systems do. |
Keywords
» Artificial intelligence » Gpt » Reinforcement learning from human feedback » Rlhf
