TAPR: Enhancing LLM Performance with a Task-Aware Prompt Rewriter

Many people find it hard to get the best answers from large language models because crafting an effective prompt takes practice and intuition. When a prompt is vague or off‑target, the model’s response can miss the point, leaving users frustrated. Researchers have introduced a system that automatically reshapes a user’s input into a version that is tuned for the specific task at hand, aiming to close that gap without requiring the user to become a prompt‑engineering expert.

What You Need to Know

The core contribution is a model called the Task‑Aware Prompt Rewriter (TAPR). TAPR takes a raw user prompt and rewrites it into a new prompt that is optimized for the downstream task the user wants the LLM to perform. The rewriting process is guided by reinforcement learning, specifically a variant known as Group Relative Policy Optimization (GRPO). In this setup, the rewriting policy receives a reward signal that reflects how well the rewritten prompt leads to a good answer.

The reward itself comes from an LLM‑acting‑as‑judge. After TAPR produces a candidate prompt, the target LLM generates an answer to the task. A separate judge LLM evaluates both the quality of the rewritten prompt (e.g., clarity, relevance) and the correctness or usefulness of the answer. The combined feedback shapes the GRPO updates, encouraging TAPR to discover rewrites that consistently improve performance across different task types.

Why It Matters

Non‑expert users often struggle to elicit high‑quality outputs from LLMs because they lack intuition about which phrasing helps the model focus. By automating prompt improvement, TAPR lowers the barrier to effective LLM use, making the technology more accessible in everyday workflows such as customer support, content creation, or educational tutoring.

From a research perspective, the approach demonstrates that a lightweight rewriting model can be trained end‑to‑end using only the target LLM and a judge LLM, without needing large amounts of human‑annotated prompt‑answer pairs. This suggests a scalable path for adapting LLMs to new domains where prompt engineering expertise is scarce.

Key Details

• Training objective: GRPO optimizes the rewriting policy by comparing each candidate rewrite against a group of alternatives and assigning rewards based on LLM‑as‑judge scores.
• Reward components: (1) Prompt quality score (clarity, task relevance) from the judge LLM; (2) Task performance score (accuracy, ROUGE, etc.) of the target LLM’s output on the rewritten prompt.
• Evaluation tasks: Question answering (NaturalQuestions), summarization (CNN/DailyMail), and arithmetic reasoning (GSM8K).
• Results: Across all three benchmarks, TAPR‑rewritten prompts yielded measurable improvements over the base model’s direct prompting—e.g., +2.3 % exact match on QA, +1.8 % ROUGE‑L on summarization, and +3.1 % accuracy on arithmetic reasoning.
• Ablation: Removing either the prompt‑quality or the task‑performance term from the reward reduced gains, indicating both signals are necessary for stable learning.
• Model size: TAPR was trained with a 7‑parameter‑count transformer, showing that a relatively small rewritter can deliver consistent benefits.

What’s Next

Future work could explore extending TAPR to multilingual settings, where the judge LLM must evaluate prompts and answers in different languages, or to interactive scenarios where the user can provide feedback on the rewritten prompt in real time. Another direction is to investigate whether the rewritter can learn to produce prompts that not only improve immediate task performance but also encourage better reasoning steps, potentially reducing the need for separate chain‑of‑thought prompting.

📌 Source: Arxiv Ai