Scholarlink

We propose an approach where human-designed logic programs guide the systematic exploration of concepts and statements using Large Language Models (LLMs). Undefined facts within these programs, (seen as abducibles), are expanded into Prolog clauses by extracting symbolic knowledge stored within an LLM’s internal representations. This collaborative framework combining human expertise and LLM-generated insights allows thorough exploration of the assumptions and potential consequences that support or challenge specific viewpoints or decisions. We demonstrate the effectiveness of this methodology by analyzing scientific theories, emerging technologies, and key policy decisions.

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

Focusing Recursive LLM Descents with Plans Expressed as Logic Programs

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Paul Tarau

摘要

We propose an approach where human-designed logic programs guide the systematic exploration of concepts and statements using Large Language Models (LLMs). Undefined facts within these programs, (seen as abducibles), are expanded into Prolog clauses by extracting symbolic knowledge stored within an LLM’s internal representations. This collaborative framework combining human expertise and LLM-generated insights allows thorough exploration of the assumptions and potential consequences that support or challenge specific viewpoints or decisions. We demonstrate the effectiveness of this methodology by analyzing scientific theories, emerging technologies, and key policy decisions.