Context-aware Prompt Tuning: Enhancing Few-Shot Learning via Optimized Context Embeddings

TMLR Paper5219 Authors

Abstract:

Large Language Models (LLMs) can perform few-shot learning using either In-Context Learning (ICL) or optimization-based methods.
While ICL typically excels in low-data regimes, optimization-based methods tend to perform better when more data is available.
This contrast raises an important question: Why do optimization-based methods struggle in low-data scenarios, and how can they be effectively combined with ICL to enhance few-shot learning? In this work, we identify overfitting as the primary limitation of optimization-based methods in few-shot learning. To address this, we propose Context-Aware Prompt Tuning (CPT), which extends ICL through a carefully designed optimization process specifically crafted to mitigate overfitting. CPT extracts richer insights from limited data while preserving the integrity of the original input samples. We validate our approach across diverse classification and question answering tasks and multiple LLM architectures. CPT consistently outperforms existing baselines across tasks and models, significantly reducing overfitting and improving generalization in few-shot scenarios.

Submission Length: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Jose_Dolz1

Submission Number: 5219