Scholarlink

<p>Retrieval-augmented generation (RAG) plays a crucial role in enhancing the capabilities of large language models (LLMs). However, traditional RAG is restricted by the temporal validity and narrow coverage of knowledge encoded in model parameters, which makes the models susceptible to hallucinations. In addition, the incorporation of redundant texts may introduce noise, thereby impairing the overall quality of the generated answers. We present a retrieval-augmented generation framework called intrinsic knowledge-aware and learning-based filtering for enhancing retrieval-augmented generation (IKF-RAG). The framework strengthens the reasoning capability of LLMs by dynamically incorporating external knowledge and is composed of three key modules, namely (1) intrinsic knowledge discrimination, which assesses whether the model can address a query based on its internal knowledge, (2) external knowledge filtering, which leverages the contextual cross-mutual information method to refine the quality and relevance of retrieved documents and (3) knowledge-adaptive generation, which produces the final answer by utilizing high-confidence contextual information. Experimental findings indicate that IKF-RAG yields notable gains in answer generation accuracy, with improvements of 7.1% on Natural Questions and 12.5% on HotpotQA over the RAG framework under the InstructGPT (text-davinci-003) setting, highlighting the framework’s effectiveness in strengthening the reasoning capability of LLMs.</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

IKF-RAG:intrinsic knowledge-aware and learning-based filtering for enhancing retrieval-augmented generation

  • 期刊论文

  • 发表日期：

    2026年2月25日

查看全文
  • Shuhuan Yan,
  • Wei Wang,
  • Hongwei Chen,
  • Jie Wang

摘要

Retrieval-augmented generation (RAG) plays a crucial role in enhancing the capabilities of large language models (LLMs). However, traditional RAG is restricted by the temporal validity and narrow coverage of knowledge encoded in model parameters, which makes the models susceptible to hallucinations. In addition, the incorporation of redundant texts may introduce noise, thereby impairing the overall quality of the generated answers. We present a retrieval-augmented generation framework called intrinsic knowledge-aware and learning-based filtering for enhancing retrieval-augmented generation (IKF-RAG). The framework strengthens the reasoning capability of LLMs by dynamically incorporating external knowledge and is composed of three key modules, namely (1) intrinsic knowledge discrimination, which assesses whether the model can address a query based on its internal knowledge, (2) external knowledge filtering, which leverages the contextual cross-mutual information method to refine the quality and relevance of retrieved documents and (3) knowledge-adaptive generation, which produces the final answer by utilizing high-confidence contextual information. Experimental findings indicate that IKF-RAG yields notable gains in answer generation accuracy, with improvements of 7.1% on Natural Questions and 12.5% on HotpotQA over the RAG framework under the InstructGPT (text-davinci-003) setting, highlighting the framework’s effectiveness in strengthening the reasoning capability of LLMs.