Scholarlink

<p>Current large language models, due to their inherent reliance on statistical correlations, are primarily optimized for generating coherent text rather than understanding deep semantic meanings, which hinders their effectiveness in solving complex reasoning tasks. Although popular prompting methods such as Chain-of-Thought (CoT) have been investigated to enhance the performance of LLMs for reasoning tasks, the intermediate reasoning steps generated remain uncontrolled, making it difficult to trace and treat hallucinations. Thus, we propose <i>VisFIT</i>, a Visual analysis method inspired by analogical transfer from Family Inheritance scenarios, utilizing Tree-structured prompts. <i>VisFIT</i> creatively draws an analogy to the dual rules of inheritance. On one hand, LLM assigns a “preference level” to each child node while performing automated reasoning based on the hierarchical prompts, which enables it to execute “testamentary inheritance". On the other hand, Users can act as “judges", transparently monitoring and modifying the reasoning process through interactive interventions, thereby implementing “statutory inheritance" as a form of guided adjustment. Specifically, <i>VisFIT</i> first interactively constructs a tree structure named <i>FIT</i> from the underlying logic in the relevant text, using it as a structural prompt for LLMs to tackle complex reasoning tasks. Visualizations at varying levels of granularity enable the tracing back of erroneous reasoning steps to their sources. Then, users can then iteratively adjust <i>FIT</i> to refine the prompts, thereby enhancing the model’s performance in specific domains. Additionally, we conducted quantitative study, case analysis, and expert review to comprehensively evaluate <i>VisFIT</i>. The results show that it improves the average accuracy by 30.37% compared to CoT across the three complex reasoning cases, providing a novel and explainable paradigm for human–LLM interaction.</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

VisFIT: a visual analysis method for complex reasoning in LLMs inspired by family inheritance scenario

  • 期刊论文

  • 发表日期：

    2026年7月1日

查看全文
  • Songyue Li,
  • Yi Ding,
  • Yize Li,
  • Mengqi Huang,
  • Xiangyang Wu,
  • Wei Xu,
  • Yongheng Wang,
  • Jie Xu,
  • Zhiguang Zhou

摘要

Current large language models, due to their inherent reliance on statistical correlations, are primarily optimized for generating coherent text rather than understanding deep semantic meanings, which hinders their effectiveness in solving complex reasoning tasks. Although popular prompting methods such as Chain-of-Thought (CoT) have been investigated to enhance the performance of LLMs for reasoning tasks, the intermediate reasoning steps generated remain uncontrolled, making it difficult to trace and treat hallucinations. Thus, we propose VisFIT, a Visual analysis method inspired by analogical transfer from Family Inheritance scenarios, utilizing Tree-structured prompts. VisFIT creatively draws an analogy to the dual rules of inheritance. On one hand, LLM assigns a “preference level” to each child node while performing automated reasoning based on the hierarchical prompts, which enables it to execute “testamentary inheritance". On the other hand, Users can act as “judges", transparently monitoring and modifying the reasoning process through interactive interventions, thereby implementing “statutory inheritance" as a form of guided adjustment. Specifically, VisFIT first interactively constructs a tree structure named FIT from the underlying logic in the relevant text, using it as a structural prompt for LLMs to tackle complex reasoning tasks. Visualizations at varying levels of granularity enable the tracing back of erroneous reasoning steps to their sources. Then, users can then iteratively adjust FIT to refine the prompts, thereby enhancing the model’s performance in specific domains. Additionally, we conducted quantitative study, case analysis, and expert review to comprehensively evaluate VisFIT. The results show that it improves the average accuracy by 30.37% compared to CoT across the three complex reasoning cases, providing a novel and explainable paradigm for human–LLM interaction.