Authors: Jack Lu, Ryan Teehan, Zhenbang Yang, and Mengye Ren

Abstract: We introduce Context Tuning, a simple and effective method to significantly enhance few-shot adaptation of large language models (LLMs) without weight updates. In-Context Learning (ICL) forms a memory representation of the demonstrations in a single forward pass but cannot refine it when insufficient. Prompt-based methods offer lightweight adaptation by optimizing a trainable prompt or prefix but initialize it independently of the demonstrations. In contrast, Context Tuning leverages the model's inherent ICL ability to initialize a trainable memory representation from demonstrations, then refines it through gradient-based optimization. Extensive evaluations on benchmarks such as CrossFit, UnifiedQA, MMLU, BIG-Bench Hard, and ARC demonstrate that Context Tuning outperforms both ICL and traditional prompt-based adaptation methods while achieving competitive accuracy with Test-Time Training at significantly higher training efficiency.

Published: 2026-07-03

Venue: The 43rd International Conference on Machine Learning (ICML 2026)