Learning to Rewrite Prompts for Bootstrapping LLMs on Downstream Tasks

TL;DR

This work tackles the problem that prompt prompts for LLMs often rely on fixed instructions, but for many NLG tasks the input content drives performance. It introduces ROI, a lightweight pipeline that rewrites the input component using a small rewriting model learned via back-translation, supplemented by a similarity-based filter to remove noisy rewrites, all without updating the LLM. The approach demonstrates consistent improvements across multi-domain machine translation, summarization, and GLUE benchmarks, highlighting gains in BLEU, RougeL, Edit Rate, accuracy, and F1 while reducing the need for expensive LLM fine-tuning. The methods offer a practical, training-efficient route to better prompt alignment with LLM preferences, with potential applicability to a broad range of downstream tasks beyond translation and sentiment analysis.

Abstract

In recent years, the growing interest in Large Language Models (LLMs) has significantly advanced prompt engineering, transitioning from manual design to model-based optimization. Prompts for LLMs generally comprise two components: the \textit{instruction}, which defines the task or objective, and the \textit{input}, which is tailored to the instruction type. In natural language generation (NLG) tasks such as machine translation, the \textit{input} component is particularly critical, while the \textit{instruction} component tends to be concise. Existing prompt engineering methods primarily focus on optimizing the \textit{instruction} component for general tasks, often requiring large-parameter LLMs as auxiliary tools. However, these approaches exhibit limited applicability for tasks like machine translation, where the \textit{input} component plays a more pivotal role. To address this limitation, this paper introduces a novel prompt optimization method specifically designed for machine translation tasks. The proposed approach employs a small-parameter model trained using a back-translation-based strategy, significantly reducing training overhead for single-task optimization while delivering highly effective performance. With certain adaptations, this method can also be extended to other downstream tasks.

Figures (4)

• Figure 1: The example on the left illustrates the sensitivity of LLMs to input data. In the context of the same translation task, providing more detailed input (bottom) leads to better results. The right figure shows the performance improvement of the LLMs on various translation datasets after rewriting.
• Figure 2: The pipeline of our proposed method for boostrapping Large Language Models. 1) The raw data is first input to the LLMs to construct the rewriting data. The generated rewritten data is then used to train the rewriting model. 2) The filtering process is applied to retain only the rewritten data that demonstrates improved performance, while the remaining data continues to utilize the original data. 3) During testing, the original data is first input to the rewriting model to obtain rewritten sentences. These rewritten sentences are subsequently input to the LLMs to generate the final results.
• Figure 3: Results under different filtering metrics. We use three thresholds: $0.3$, $0.5$ and $0.8$.
• Figure 4: Translation performance when training a rewriting model using 100, 1000, 5000, and 10000 pieces of rewritten data, respectively.