Mitigating Copy Bias in In-Context Learning through Neuron Pruning

Understand

Large language models (LLMs) have demonstrated impressive few-shot in-context learning (ICL) abilities.

• Still, we show that they are sometimes prone to a `copying bias', where they copy answers from provided examples instead of learning the underlying patterns.
• In this work, we propose a novel and simple method to mitigate such copying bias.
• First, we create a synthetic task and use the Integrated Gradients method to identify neurons that prioritize copying over generalization.