Scholarlink

<p>Advancements in large language models have highlighted the critical role of prompt quality in achieving accurate responses. However, existing research on prompt optimization, mainly based on exploratory methods like reinforcement learning, suffers from challenges such as the vast search space of prompts, limited flexibility, and poor adaptability. To tackle these challenges, we introduce the <i>weak-to-strong generalization</i> concept into prompt optimization for the first time and propose the weak-to-strong prompt correction (W2S) framework. The core principle of W2S is to simplify the optimization process by starting with weak prompts and progressively correcting them into stronger prompts. W2S begins by constructing a weak-to-strong prompt dataset using LLM, then employs a two-stage weak-to-strong prompt correction: supervised fine-tuning to learn transfer patterns from weak prompts to strong prompts and direct preference optimization to mitigate potential errors. Experimental results demonstrate that W2S significantly outperforms existing methods, achieving superior adaptability across diverse generative tasks and offering an incremental optimization mechanism for prompt quality enhancement. The code is available at <a href="https://github.com/lrgao/W2S">https://github.com/lrgao/W2S</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

W2S: Weak-to-Strong Prompt Correction for Large Language Models

  • 期刊论文

  • 发表日期：

    2026年3月2日

查看全文
  • Lirong Gao,
  • Xinyi Wang,
  • Hao Chen,
  • Ru Peng,
  • Qi Zhang,
  • Yiming Zhang,
  • Wentao Ye,
  • Haoze Li,
  • Haobo Wang,
  • Junbo Zhao

摘要

Advancements in large language models have highlighted the critical role of prompt quality in achieving accurate responses. However, existing research on prompt optimization, mainly based on exploratory methods like reinforcement learning, suffers from challenges such as the vast search space of prompts, limited flexibility, and poor adaptability. To tackle these challenges, we introduce the weak-to-strong generalization concept into prompt optimization for the first time and propose the weak-to-strong prompt correction (W2S) framework. The core principle of W2S is to simplify the optimization process by starting with weak prompts and progressively correcting them into stronger prompts. W2S begins by constructing a weak-to-strong prompt dataset using LLM, then employs a two-stage weak-to-strong prompt correction: supervised fine-tuning to learn transfer patterns from weak prompts to strong prompts and direct preference optimization to mitigate potential errors. Experimental results demonstrate that W2S significantly outperforms existing methods, achieving superior adaptability across diverse generative tasks and offering an incremental optimization mechanism for prompt quality enhancement. The code is available at https://github.com/lrgao/W2S.