Scholarlink

As AI models become central to societal functions, they inevitably attract the attention of malicious actors, resulting in the proliferation of counterfeit AI models. Therefore, identity verification, similar to user recognition, is essential for AI models. However, owing to probabilistic variability in outputs, particularly in large language models (LLMs), and continuous ability enhancement through autonomous learning, the typical user recognition method of enrolling and presenting the same information cannot be directly applied. To address these issues, this study proposes identifying AI models by “assessing the abilities of AI models.” The identification method based on this concept is called the Completely Automated Public Test to Tell Ability of Artificial Intelligence (CAPT-AI). To examine the feasibility of our proposed method, we conduct a preliminary study focusing on LLMs, serving as a foundational experiment for CAPT-AI.

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

A Study on AI Model Identification Based on Ability Differences: An Initial Approach Using Semantic Understanding Ability

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Seiya Kajihara,
  • Ryunosuke Harada,
  • Toma Ogiri,
  • Wataru Hatakeyama,
  • Kazuki Takabayashi,
  • Seiji Sato,
  • Nami Ashizawa,
  • Toshiki Shibahara,
  • Osamu Saisho,
  • Tetsushi Ohki,
  • Masakatsu Nishigaki

摘要

As AI models become central to societal functions, they inevitably attract the attention of malicious actors, resulting in the proliferation of counterfeit AI models. Therefore, identity verification, similar to user recognition, is essential for AI models. However, owing to probabilistic variability in outputs, particularly in large language models (LLMs), and continuous ability enhancement through autonomous learning, the typical user recognition method of enrolling and presenting the same information cannot be directly applied. To address these issues, this study proposes identifying AI models by “assessing the abilities of AI models.” The identification method based on this concept is called the Completely Automated Public Test to Tell Ability of Artificial Intelligence (CAPT-AI). To examine the feasibility of our proposed method, we conduct a preliminary study focusing on LLMs, serving as a foundational experiment for CAPT-AI.