Selecting Demonstrations for Many-Shot In-Context Learning via Gradient Matching

TL;DR

This paper introduces a gradient matching method for selecting demonstrations in many-shot in-context learning, significantly improving performance over random selection across various large language models and datasets.

Contribution

It proposes a novel gradient matching technique for demonstration selection that aligns fine-tuning gradients, enhancing in-context learning effectiveness without additional training.

Findings

Outperforms random selection on large language models from 4-shot to 128-shot.

Achieves 4% improvement on Qwen2.5-72B and Llama3-70B.

Improves performance on 9 diverse datasets.

Abstract

In-Context Learning (ICL) empowers Large Language Models (LLMs) for rapid task adaptation without Fine-Tuning (FT), but its reliance on demonstration selection remains a critical challenge. While many-shot ICL shows promising performance through scaled demonstrations, the selection method for many-shot demonstrations remains limited to random selection in existing work. Since the conventional instance-level retrieval is not suitable for many-shot scenarios, we hypothesize that the data requirements for in-context learning and fine-tuning are analogous. To this end, we introduce a novel gradient matching approach that selects demonstrations by aligning fine-tuning gradients between the entire training set of the target task and the selected examples, so as to approach the learning effect on the entire training set within the selected examples. Through gradient matching on relatively small models, e.g., Qwen2.5-3B or Llama3-8B, our method consistently outperforms random selection on larger LLMs from 4-shot to 128-shot scenarios across 9 diverse datasets. For instance, it surpasses random selection by 4% on Qwen2.5-72B and Llama3-70B, and by around 2% on 5 closed-source LLMs. This work unlocks more reliable and effective many-shot ICL, paving the way for its broader application.