PACIT: Unlocking the Power of Examples for Better In-Context Instruction Tuning

TL;DR

PACIT is a novel in-context instruction tuning method inspired by pedagogical principles, which improves large language models' performance by encouraging active learning from examples, outperforming previous methods on various tasks.

Contribution

Introducing PACIT, a simple yet effective in-context instruction tuning approach that leverages desirable difficulty to enhance model learning from positive and negative examples.

Findings

PACIT outperforms ICIT baseline by up to 9.16 ROUGE-L points.

PACIT improves performance on both in-domain and out-domain tasks.

Self-instruct generated examples also benefit from PACIT.

Abstract

Instruction tuning enhances the instruction following ability of large language models by finetuning with supervised instruction data. Previous work proposes in-context instruction tuning (ICIT) where specific positive or negative examples are incorporated into the prompt for better performance. In this work, we propose PACIT, a simple and effective in-context instruction tuning method, inspired by the pedagogical concept of desirable difficulty. The PACIT method unlocks the power of examples by encouraging the model to actively learn to grasp the distinctions between the positive and negative examples instead of merely reading. The model is expected to first verify the correctness of the provided example according to the task description, which is then set as the condition for generating a better response to the task instance. Our extensive experiments prove the effectiveness of PACIT, outperforming ICIT baseline on both in-domain and out-domain tasks up to 9.16 and 3.14 average ROUGE-L scores, respectively. Moreover, PACIT can notably enhance the performance of instruction tuning even when all positive and negative examples are generated with a self-instruct method.