IOS Press Ebooks - Adaptive Instruction Induction for Enhancing Large Language Model Performance (original) (raw)

Abstract

Instructions, as the primary means of using large language models (LLMs), significantly impact the results. Automatically inducing instructions from few-shot instances is meaningful, yet the induced instructions suffer from two inherent divergences: (1) systematic discrepancies across tasks, and (2) instance-level heterogeneity within a task. To address these challenges, inspired by human analogical reasoning, we propose AdaIn, an iteratively adaptive instruction induction framework that leverages instance-level structural similarities. AdaIn groups instances by data features to induce tailored instructions, and adaptively applies them to new samples. Instruction performance is further utilized to guide updates„ enabling iterative identification and refinement of ineffective instructions. We conducted experiments on different tasks to verify our method and the results demonstrate that it outperforms the SOTA results. Ablation experiments indicate that the adaptive strategy in induction and selection instruction contributes much to performance.