Calibrate Before Use: Improving Few-Shot Performance of Language Models

Tony Z. Zhao, Eric Wallace, Shi Feng, Dan Klein, Sameer Singh
Paper
Github

Tags: [prompt_engineering, llm]

Abstract

• Shows that few-shot prompting for large language models like GPT-3 can be unstable

• Demonstrate that this instability arises from the bias of language models towards predicting certain answers

• Mitigate model biases by first estimating the model’s bias towards each answer by asking for its prediction when given the training prompt and a content-free test input (e.g. “N/A”)

• Then fit calibration parameters that cause the prediction for this input to be uniform across answers

Introduction

Few-shot learning

• Ability to learn tasks with limited examples

• Few-shot “in-context” learning shows that model can learn without parameter updates

• Example:

Input: Subpar acting.  Sentiment: Negative
Input: Beautiful film. Sentiment: Positive
Input: Amazing.        Sentiment:

Instability

• Shows GPT-3 accuracy can be highly unstable across different prompts

• Causes of instability:

  • Majority label bias

  • Recency bias

  • Common token bias