Summary of Mitigating Selection Bias with Node Pruning and Auxiliary Options, by Hyeong Kyu Choi et al.
Mitigating Selection Bias with Node Pruning and Auxiliary Options
by Hyeong Kyu Choi, Weijie Xu, Chi Xue, Stephanie Eckman, Chandan K. Reddy
First submitted to arxiv on: 27 Sep 2024
Categories
- Main: Artificial Intelligence (cs.AI)
- Secondary: None
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 research paper explores the issue of selection bias in large language models (LLMs) when responding to multiple-choice questions. Unwarranted preferences for certain options can significantly impact reliability in automated systems. Previous solutions have used debiasing methods, but this work focuses on the internal representation of selection bias within LLMs. The authors introduce two novel approaches: Bias Node Pruning (BNP), which eliminates linear layer parameters contributing to bias, and Auxiliary Option Injection (AOI), a simple input modification technique for debiasing compatible with black-box LLMs. Additionally, they propose Choice Kullback-Leibler Divergence (CKLD) as a more systematic evaluation metric for selection bias, addressing the insensitivity of existing metrics to label imbalance. The authors demonstrate that their methods are robust and adaptable across various datasets when applied to three LLMs. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This research helps make sure that computer programs that answer questions don’t have a favorite answer just because it’s easy or popular. This problem, called selection bias, can make the results unreliable. The authors of this paper found two new ways to fix this issue: one way is to get rid of some parts inside the computer program that cause the bias, and another way is to modify what goes into the program before it gives an answer. They also came up with a better way to measure how well these fixes work. The authors tested their methods on several types of language models and found that they were effective in reducing selection bias. |
Keywords
» Artificial intelligence » Pruning
