Scholarlink

<p>In system fault diagnosis tasks, LLM can identify system errors as inputs and generate diagnostic results as outputs. However, while system monitoring data is abundant as input, high-quality annotated output data remains scarce, typically requiring manual labeling and validation. This asymmetry between input and output presents significant challenges for model training. To tackle this challenge, this paper proposes a method combining Reverse Verification Chain of Thought with Value-Driven Direct Preference Optimization. This approach injects a local reverse verification chain mechanism into the LLM through manually labeled data and introduces a score model to evaluate the generated reasoning chains. Based on these scores, secondary reinforcement learning fine-tuning is conducted. This fine-tuning strategy not only significantly enhances the model reasoning ability in complex fault scenarios but also greatly reduces the reliance on large amounts of manually labeled data. Experimental results show that it significantly improves the LLM reasoning ability at each fine-tuning stage, demonstrating stronger accuracy and interpretability compared to baseline methods. This research provides new insights and methodologies for the automated fault diagnosis of complex systems, advancing the development of fault diagnosis technologies and showing broad potential for practical applications in industries such as industrial systems, cloud computing, and software development.</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

Local self-verified chain-of-thought and preference alignment: a two-stage ReVCoT-DPO fine-tuning framework for system log fault diagnosis

  • 期刊论文

  • 发表日期：

    2026年4月3日

查看全文
  • Zhichao Zhen,
  • Chen Ling,
  • Lizhu Mi,
  • Hongbin Zhang

摘要

In system fault diagnosis tasks, LLM can identify system errors as inputs and generate diagnostic results as outputs. However, while system monitoring data is abundant as input, high-quality annotated output data remains scarce, typically requiring manual labeling and validation. This asymmetry between input and output presents significant challenges for model training. To tackle this challenge, this paper proposes a method combining Reverse Verification Chain of Thought with Value-Driven Direct Preference Optimization. This approach injects a local reverse verification chain mechanism into the LLM through manually labeled data and introduces a score model to evaluate the generated reasoning chains. Based on these scores, secondary reinforcement learning fine-tuning is conducted. This fine-tuning strategy not only significantly enhances the model reasoning ability in complex fault scenarios but also greatly reduces the reliance on large amounts of manually labeled data. Experimental results show that it significantly improves the LLM reasoning ability at each fine-tuning stage, demonstrating stronger accuracy and interpretability compared to baseline methods. This research provides new insights and methodologies for the automated fault diagnosis of complex systems, advancing the development of fault diagnosis technologies and showing broad potential for practical applications in industries such as industrial systems, cloud computing, and software development.