TAPR: Task-Aware Prompt Rewriter Boosts LLM Performance via Reinforcement Learning
Researchers have unveiled a novel model named Task-Aware Prompt Rewriter (TAPR), designed to improve the efficacy of Large Language Models (LLMs) by transforming user prompts into optimized formats for specific tasks. This development is outlined in a paper on arXiv (ID: 2607.28657). TAPR tackles the frequent issue faced by non-expert users who find it challenging to create effective prompts. Utilizing reinforcement learning with Group Relative Policy Optimization (GRPO), the model receives rewards based on evaluations from LLMs assessing both the revised prompt and its task output. Testing on various tasks, including question answering, summarization, and arithmetic reasoning, demonstrates significant improvements in prompt rewriting. Fine-tuning Phi-4-mini-instruct as TAPR's base model results in clearer, more instructive prompts, enhancing accuracy on key benchmarks. The paper, titled 'TAPR: Enhancing LLM Performance with a Task-Aware Prompt Rewriter', emphasizes the promise of automated prompt optimization in broadening access to LLMs.
Key facts
- TAPR (Task-Aware Prompt Rewriter) is a model that reformulates user prompts to improve downstream LLM performance.
- TAPR is trained using reinforcement learning with Group Relative Policy Optimization (GRPO).
- Rewards are derived from LLM-as-judge evaluations of the reformulated prompt and the task output.
- Experiments were conducted on question answering, summarization, and arithmetic reasoning tasks.
- Fine-tuning Phi-4-mini-instruct as the base model for TAPR yields clearer and more instructive prompts.
- The method shows consistent gains over base models in prompt rewriting ability.
- The paper is available on arXiv with ID 2607.28657.
- The work aims to help non-expert users by automating prompt optimization.
Entities
Institutions
- arXiv