Scholarlink

Selecting appropriate examples for in-context learning significantly impacts the performance of Large Language Models. In this paper, we show that leveraging LLMs’ positional biases and incorporating knowledge of class distribution can improve classification outcomes, especially for underrepresented classes. We introduce Class-few-shot, a method that balances class representation among few-shot examples. To investigate this, we conduct almost 10,000 experiments on four datasets and three models, cross-checking how different biases affect models’ performance and how they interact. We show that presenting classes from the most to least numerous using an alternating pattern leads to better results than standard few-shot prompting with the same number of examples. Additionally, we compare the general few-shot and Class-few-shot results, outlining the strengths of both approaches. All of our raw experiment results, prompts and codes are publicly available on GitHub ( https://github.com/olorules/Class-few-shot ).

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Leveraging Positional Bias of LLM In-Context Learning with Class-Few-Shot and Maj-Min Alternating Ordering

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Aleksander Szczęsny,
  • Maciej Markiewicz,
  • Łukasz Radliński,
  • Przemysław Kazienko

摘要

Selecting appropriate examples for in-context learning significantly impacts the performance of Large Language Models. In this paper, we show that leveraging LLMs’ positional biases and incorporating knowledge of class distribution can improve classification outcomes, especially for underrepresented classes. We introduce Class-few-shot, a method that balances class representation among few-shot examples. To investigate this, we conduct almost 10,000 experiments on four datasets and three models, cross-checking how different biases affect models’ performance and how they interact. We show that presenting classes from the most to least numerous using an alternating pattern leads to better results than standard few-shot prompting with the same number of examples. Additionally, we compare the general few-shot and Class-few-shot results, outlining the strengths of both approaches. All of our raw experiment results, prompts and codes are publicly available on GitHub ( https://github.com/olorules/Class-few-shot ).