Introduction

Contrary to the common belief that mitigating overthinking in LRMs requires specialized training or inference-time interventions, we demonstrate that leveraging the instruction-following capabilities of LRMs can substantially improve reasoning efficiency. By treating both LRMs and closed-source APIs as black-box communicators, we introduce Whisper, an iterative refinement framework that generates high-quality persuasive prompts from diverse perspectives, to elicit concise responses while maintaining reasoning performance.

We explore five distinct types of persuasive prompts, including emotional appeal, threat, evidence-based persuasion, role-playing, and detailed instructions. Experiments demonstrate that Whisper consistently reduces token usage while preserving performance. Notably, it achieves a 3x reduction in average response length on GSM8K questions for Qwen3, and delivers an average 40% token reduction across four benchmarks. For closed-source APIs, Whisper effectively reduces a 2x token usage on MATH-500 for Claude-3.7 and Gemini-2.5.

Update

2025.10.14: We have released the evaluation scripts and top-performing prompts in Whisper. Check it out!

Todo

• Release instructions and scripts for prompt candidate evaluation

Installation

conda create whisper python=3.10
conda activate whisper
cd Whisper
uv pip install -r requirements.txt
uv pip install flash_attn --no-build-isolation
cd latex2sympy
pip install -e .

Evaluation

Modify and run command lines in sh/qwen3/eval_qwen3.sh, the results will be stored in output/.

bash sh/qwen3/eval_qwen3.sh

Acknowledgments

This codebase is built from Qwen2.5-Math.

Citation

If you find the resources in this repository useful, please cite our paper:

@misc{xia2025whisper,
      title={Merlin's Whisper: Enabling Efficient Reasoning in Large Language Models via Black-box Persuasive Prompting}, 
      author={Heming Xia and Cunxiao Du and Rui Li and Chak Tou Leong and Yongqi Li and Wenjie Li},
      year={2025},
      eprint={2510.10528},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2510.10528}, 
}