TAPR Rewrites Prompts to Boost LLM Performance

Achieving optimal performance from large language models (LLMs) often necessitates expertly crafted prompts, a skill that can be a significant barrier for non-specialist users. This research addresses this challenge by presenting the Task-Aware Prompt Rewriter (TAPR), a novel model designed to automatically transform user-provided prompts into versions that are specifically optimized for a given task, thereby boosting the performance of the target LLM. TAPR is trained using a reinforcement learning approach, specifically Group Relative Policy Optimization (GRPO). The reward signals for this training are derived from an "LLM-as-judge" evaluation system, which assesses both the quality of the reformulated prompt and the accuracy of the corresponding output generated by the downstream LLM. This iterative feedback loop allows TAPR to learn how to generate clearer and more instructive language in its rewritten prompts. Experimental results across a variety of tasks, including question answering, summarization, and arithmetic reasoning, consistently demonstrated that TAPR yields significant performance gains over baseline models. Fine-tuning a model like Phi-4-mini-instruct as the TAPR base model resulted in prompts that led to higher accuracy on established benchmarks such as Natural Questions and GSM8K. The code for TAPR has been made publicly available to support further research and reproducibility.