Rectifying Demonstration Shortcut in In-Context Learning

TL;DR

This paper identifies the 'Demonstration Shortcut' in large language models' in-context learning, where models rely on priors rather than input-label relationships, and proposes In-Context Calibration to address this issue, improving learning across multiple models and settings.

Contribution

The paper introduces In-Context Calibration, a novel demonstration-aware calibration method that rectifies the Demonstration Shortcut in LLMs, enhancing their ability to learn input-label relationships.

Findings

In-Context Calibration significantly improves ICL performance.

Method generalizes across multiple LLM families.

Effective in both standard and task learning settings.

Abstract

Large language models (LLMs) are able to solve various tasks with only a few demonstrations utilizing their in-context learning (ICL) abilities. However, LLMs often rely on their pre-trained semantic priors of demonstrations rather than on the input-label relationships to proceed with ICL prediction. In this work, we term this phenomenon as the 'Demonstration Shortcut'. While previous works have primarily focused on improving ICL prediction results for predefined tasks, we aim to rectify the Demonstration Shortcut, thereby enabling the LLM to effectively learn new input-label relationships from demonstrations. To achieve this, we introduce In-Context Calibration, a demonstration-aware calibration method. We evaluate the effectiveness of the proposed method in two settings: (1) the Original ICL Task using the standard label space and (2) the Task Learning setting, where the label space is replaced with semantically unrelated tokens. In both settings, In-Context Calibration demonstrates substantial improvements, with results generalized across three LLM families (OPT, GPT, and Llama2) under various configurations.