Context-aware Prompt Tuning: Advancing In-Context Learning with Adversarial Methods

ACL ARR 2025 February Submission61 Authors

Abstract:

Large Language Models (LLMs) can perform few-shot learning using In-Context Learning (ICL) or optimization-based methods.
ICL is more effective in low-data regimes, while optimization-based methods excel with larger datasets. This contrast raises a key question: why optimization-based methods face challenges in low-data regimes, and how can these methods be effectively integrated with ICL to enhance few-shot learning? In this work, we identify overfitting as the primary limitation of optimization-based methods in few-shot settings and introduce Context-aware Prompt Tuning (CPT), a method that combines the strengths of ICL, Prompt Tuning (PT), and adversarial techniques. CPT initializes the context with training examples, similar to ICL, and then applies an optimization process inspired by PT and adversarial techniques. Through iterative adaptation, CPT effectively balances flexibility and stability, allowing it to derive deeper insights from limited data while preserving the integrity of input samples. Our method achieves superior accuracy across multiple classification tasks and LLM architectures, consistently outperforming existing baselines and effectively mitigating overfitting challenges in few-shot scenarios.

Paper Type: Long

Research Area: Machine Learning for NLP

Research Area Keywords: Efficient/Low-Resource Methods for NLP, Generation, Machine Learning for NLP

Contribution Types: NLP engineering experiment, Approaches to low-resource settings

Languages Studied: In-Context Learning,

Submission Number: 61