Scholarlink

With the emergence of multimodal large models, the problem of hallucination has been plaguing their development and deployment. How to reliably detect the presence of hallucinations in mLLMs has become an important issue. We propose UHDF, which replaces the closed-source models in it with open-source models by improving UniHD [15], and dramatically outperforms it. By optimizing the external information used in UniHD and achieving decoupling between different external information sources, we minimize the hallucinations introduced in pipeline, and thus improve the effectiveness of hallucinations detection. UHDF using the open-source model outperforms UniHD using the closed-source model (GPT-4v), achieving 86.6% (dev set)/85.3% (test set) on MacroF1 and achieved the first place in NLPCC2024 Shared Task 10 Track1 (Open Source). Our code and models are available at https://github.com/codetalker125/UHDF .

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

UHDF: Hallucination Detection Using Open Source Models Beyond Close Source Models Methods

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Dongxu Liu,
  • Bufan Xu,
  • Zhilong Zhao,
  • Bing Xu,
  • Muyun Yang

摘要

With the emergence of multimodal large models, the problem of hallucination has been plaguing their development and deployment. How to reliably detect the presence of hallucinations in mLLMs has become an important issue. We propose UHDF, which replaces the closed-source models in it with open-source models by improving UniHD [15], and dramatically outperforms it. By optimizing the external information used in UniHD and achieving decoupling between different external information sources, we minimize the hallucinations introduced in pipeline, and thus improve the effectiveness of hallucinations detection. UHDF using the open-source model outperforms UniHD using the closed-source model (GPT-4v), achieving 86.6% (dev set)/85.3% (test set) on MacroF1 and achieved the first place in NLPCC2024 Shared Task 10 Track1 (Open Source). Our code and models are available at https://github.com/codetalker125/UHDF .