Scholarlink

In this work, we present the preliminary results on testing the self-consistency and cross-consistency of various LLMs on supplements recommendation task. The Kruskal-Wallis test is performed on pair-wise comparison among models to determine the consistency. In our experiment, we use anonymous health records of ten users and have six LLMs generate 20 supplement recommendations for each user. Each of the responses is then rated by each of the three GPT models. The results provide an insight into how consistent and reliable LLM-based assessments are.

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

Toward an Understanding of the Self-Coherence and the Cross-Coherence of LLMs An Empirical Investigation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Farouq Anbar,
  • Andrei Mikriukov,
  • Yaroslav Plaksin,
  • Vladimir Sitnikov,
  • Giancarlo Succi,
  • Alexander Tormasov,
  • Ekaterina Trofimova

摘要

In this work, we present the preliminary results on testing the self-consistency and cross-consistency of various LLMs on supplements recommendation task. The Kruskal-Wallis test is performed on pair-wise comparison among models to determine the consistency. In our experiment, we use anonymous health records of ten users and have six LLMs generate 20 supplement recommendations for each user. Each of the responses is then rated by each of the three GPT models. The results provide an insight into how consistent and reliable LLM-based assessments are.