Scholarlink

<p>Conversational recommender systems (CRS) enhance system interactivity to improve user satisfaction through multi-round conversations. Large language models (LLMs) unlock new CRS paradigms with strong natural language processing capabilities. However, existing LLM-based methods still suffer from two issues: (1) optimizing LLMs for conversational recommendation is constrained by inherent text comprehension capabilities; and (2) LLMs excel at aimless conversations but struggle with goal-directed ones, often deviating from the recommended target. To address these issues, we propose a CRS model based on LLMs named MSOCR, which generates optimizations for LLMs by evaluating and reflecting on recommendation results. This enables LLMs to deepen their cognition of the text through multi-round trial-and-error learning. Moreover, we design an act-decision framework that judges whether to recommend or continue conversing to capture user implicit preferences based on conversation history and known user preferences. Our study facilitates LLMs’ deeper understanding of CRS and provides a more effective scheme for LLMs to control conversation flow. Extensive experiments on three public CRS datasets demonstrate the effectiveness of our method compared to State-of-the-art baselines.</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

Multi-round self-optimization with large language models for conversational recommendation

  • 期刊论文

  • 发表日期：

    2026年3月18日

查看全文
  • Qinyang He,
  • Yihao Zhang,
  • Kaibei Li,
  • Xiaokang Li,
  • Xibin Wang

摘要

Conversational recommender systems (CRS) enhance system interactivity to improve user satisfaction through multi-round conversations. Large language models (LLMs) unlock new CRS paradigms with strong natural language processing capabilities. However, existing LLM-based methods still suffer from two issues: (1) optimizing LLMs for conversational recommendation is constrained by inherent text comprehension capabilities; and (2) LLMs excel at aimless conversations but struggle with goal-directed ones, often deviating from the recommended target. To address these issues, we propose a CRS model based on LLMs named MSOCR, which generates optimizations for LLMs by evaluating and reflecting on recommendation results. This enables LLMs to deepen their cognition of the text through multi-round trial-and-error learning. Moreover, we design an act-decision framework that judges whether to recommend or continue conversing to capture user implicit preferences based on conversation history and known user preferences. Our study facilitates LLMs’ deeper understanding of CRS and provides a more effective scheme for LLMs to control conversation flow. Extensive experiments on three public CRS datasets demonstrate the effectiveness of our method compared to State-of-the-art baselines.