Scholarlink

<p>This paper introduces a novel approach named MOCL-CoT (Multi-Objective Contrastive Learning with Chain-of-Thought) designed to enhance the performance of Large Language Models (LLMs) in scenarios involving complex relational structures. While LLMs have achieved remarkable success across various natural language processing tasks, their efficiency can be further improved through contrastive learning. However, in intricate contexts with multiple relationships, negative samples often fail to provide effective contrast, limiting their utility. Experimental results indicate that certain negative samples do not yield the desired effect, highlighting the need for appropriate metrics to evaluate LLM-generated negative samples. To address this issue, MOCL-CoT captures the relationships and entities within questions to assist LLMs in comprehending context. It employs Chain-of-Thought (CoT) reasoning to guide LLMs in identifying underlying relationships within questions. Furthermore, MOCL-CoT utilizes multi-objective optimization algorithms, specifically Pareto and TOPSIS, to evaluate and filter the generated relational triples. This process involves intentionally replacing correct relationships in the original questions with incorrect ones to create highly relevant negative samples. Experimental evaluations demonstrate that, compared to existing CoT prompting methods, MOCL-CoT achieves performance improvements across various reasoning tasks. Analysis reveals that MOCL-CoT enhances LLMs’ understanding of negative samples within the CoT framework, significantly boosting the accuracy of question-answering tasks and effectively enhancing the models’ reasoning capabilities.</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

Chain-of-thought reasoning enhancement through multi-objective optimized contrastive learning with negative examples

  • 期刊论文

  • 发表日期：

    2026年1月29日

查看全文
  • Hongwei Chen,
  • Jiahui Zhu,
  • Wei Wang,
  • Changyong Lin

摘要

This paper introduces a novel approach named MOCL-CoT (Multi-Objective Contrastive Learning with Chain-of-Thought) designed to enhance the performance of Large Language Models (LLMs) in scenarios involving complex relational structures. While LLMs have achieved remarkable success across various natural language processing tasks, their efficiency can be further improved through contrastive learning. However, in intricate contexts with multiple relationships, negative samples often fail to provide effective contrast, limiting their utility. Experimental results indicate that certain negative samples do not yield the desired effect, highlighting the need for appropriate metrics to evaluate LLM-generated negative samples. To address this issue, MOCL-CoT captures the relationships and entities within questions to assist LLMs in comprehending context. It employs Chain-of-Thought (CoT) reasoning to guide LLMs in identifying underlying relationships within questions. Furthermore, MOCL-CoT utilizes multi-objective optimization algorithms, specifically Pareto and TOPSIS, to evaluate and filter the generated relational triples. This process involves intentionally replacing correct relationships in the original questions with incorrect ones to create highly relevant negative samples. Experimental evaluations demonstrate that, compared to existing CoT prompting methods, MOCL-CoT achieves performance improvements across various reasoning tasks. Analysis reveals that MOCL-CoT enhances LLMs’ understanding of negative samples within the CoT framework, significantly boosting the accuracy of question-answering tasks and effectively enhancing the models’ reasoning capabilities.