Prompting with In-Context Learning

Explore the power and limitations of In-Context Learning (ICL) with Large Language Models for sentiment analysis.

This project explores the power and limitations of In-Context Learning (ICL). We designed prompts to guide a base Large Language Model (Qwen3-0.6B-Base) to perform sentiment analysis on the SST-2 dataset, without any fine-tuning.

Key topics:

  • Prompt design strategies for base LLMs
  • Few-shot and zero-shot in-context learning
  • Evaluation on SST-2 sentiment classification benchmark

Finding:

Under the zero-shot setting, the sentiment classification accuracy reached 82.5%. As more examples were added to the prompt context, the accuracy steadily increased. However, providing 16 examples (k=16) did not yield better performance compared to the k=8 setting, indicating a saturation point in the few-shot learning curve.

Timeline: Spring 2026