Summary of Anchored Answers: Unravelling Positional Bias in Gpt-2’s Multiple-choice Questions, by Ruizhe Li et al.
Anchored Answers: Unravelling Positional Bias in GPT-2’s Multiple-Choice Questions
by Ruizhe Li, Yanjun Gao
First submitted to arxiv on: 6 May 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI); 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 A recently published study investigates the Large Language Models (LLMs) GPT-4 and LLaMA families, which have shown impressive performance across various tasks, including multiple-choice questions (MCQs). However, researchers discovered that these models exhibit a positional bias, particularly an even worse anchored bias in the GPT-2 family. This bias causes the models to consistently favor the first choice ‘A’ in MCQs during inference, which challenges the integrity of their decision-making process. The study employs the mechanistic interpretability approach to identify internal modules within GPT-2 models responsible for this bias and proposes targeted interventions to mitigate it. By updating specific value vectors and recalibrating attention patterns, researchers effectively neutralize the preference for the first choice ‘A’ and improve overall MCQ prediction accuracy for the GPT-2 family across various datasets. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Large Language Models (LLMs) like GPT-4 and LLaMA have gotten really good at answering multiple-choice questions. But they have a problem: they often choose the first answer, just because it’s the first one! This is called an “anchored bias”. Researchers looked closely at how these models work to figure out where this bias comes from. They found that it’s due to special parts of the model that are responsible for making decisions. To fix this problem, they came up with a plan to update and adjust some of these special parts. This made their predictions even better! |
Keywords
» Artificial intelligence » Attention » Gpt » Inference » Llama
