Introduction to LLMs and In-Context Learning

Large Language Models (LLMs) like GPT-3 have shown impressive performance in understanding and generating language. This study focuses on their ability in in-context learning (ICL) for text classification, where LLMs make predictions based on given examples. However, these models can be biased, especially if the examples are not evenly distributed across different classes.

The Problem of Majority Label Bias

The study investigates majority label bias in LLMs. This bias occurs when most examples given to the model belong to one category, making the model more likely to predict that category. It’s a common issue in real-world settings due to data collection limitations.

Investigating LLM Robustness

Researchers tested several LLMs, including OpenLlama and Falcon-40B, to see how well they handle skewed example distributions. They used datasets with different types of questions and statements to measure the models’ performance and bias.

Key Findings

  1. Model Size Matters: Larger models tend to be more robust against majority label bias.
  2. Impact of Instructions: Adding clear task instructions to the prompts helps models perform better, especially in skewed scenarios.
  3. Varying Robustness: The robustness of LLMs to bias varies widely across different models and tasks.

Specific Observations

  • Some models showed up to 90% robustness against majority label bias.
  • Models with instructions performed better in extremely skewed distributions.
  • Larger models were generally more sensitive to the skewness in the absence of clear instructions.

Conclusion and Future Directions

The study concludes that while LLMs do show a degree of robustness to majority label bias, this varies with model size, task complexity, and the presence of instructions. Future research might focus on controlled output generation and fine-tuning LLMs specifically to handle majority label bias more effectively.

Reference

Check the original text “How Robust are LLMs to In-Context Majority Label Bias?” by Gupta, K. et. al., December 22, 2023 for your reference.