Scholarlink

<p>Large Language Models (LLMs) exhibit strong generative capabilities but remain limited in structured knowledge domains due to factual inconsistency, shallow multi-hop reasoning, and weak alignment with domain constraints. To address these limitations, this paper presents EduRAG-Compose, a unified hybrid retrieval–generation architecture that integrates hierarchical coarse-to-fine retrieval, dynamic knowledge graph traversal, cache-augmented inference, and compositional multi-step generation into a single framework for structured domain reasoning. The architecture combines clustered dense retrieval for efficient evidence selection, graph-based dependency traversal for multi-hop inference, recursive retrieval–generation cycles for compositional reasoning, and a caching layer to reduce end-to-end latency. A domain-alignment classifier and a constrained decoding strategy further support structured and interpretable output generation consistent with underlying semantic dependencies. The framework is evaluated on structured-domain benchmarks including Science Question Answering (ScienceQA), Textbook Question Answering (TQA), and interaction-derived query sets from EdNet, focusing on retrieval quality, reasoning coherence, and generative consistency. Across these datasets, EduRAG-Compose achieves 91.2 ± 1.3% factual accuracy, outperforming vanilla Retrieval-Augmented Generation (RAG) at 78.4 ± 2.1%, ColBERT-based Retrieval-Augmented Generation (ColBERT-RAG) at 82.9 ± 1.9%, and Generative Pre-trained Transformer with retrieval (GPT + RAG) at 88.5 ± 1.5%. Domain-alignment performance improves from 0.68 ± 0.05 to 0.89 ± 0.03, while structural reasoning consistency increases from 0.61 ± 0.06 to 0.87 ± 0.04, with an average response latency of 4.8 ± 0.7&#xa0;s, comparable to baseline retrieval systems and substantially lower than Chain-of-Thought (CoT) prompting. All reported improvements are statistically significant based on paired Wilcoxon signed-rank tests (<i>p</i> &lt; 0.01). Qualitative reasoning traces further demonstrate that explicit retrieval paths and graph traversal sequences enhance interpretability and support human-in-the-loop verification. Rather than introducing new retrieval or generation algorithms, the primary contribution lies in the systematic integration of complementary retrieval and reasoning mechanisms into a single transparent architecture for structured domains. Limitations related to dataset scope, computational overhead, and text-only processing are acknowledged, and future work will explore multimodal extensions, computational optimization for real-time deployment, and adaptive reasoning mechanisms capable of incorporating continuous feedback or evolving knowledge sources.</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

Hybrid retrieval generation for structured reasoning with large language models

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月10日

查看全文
  • Rathinasamy Muthusami,
  • Kandhasamy Saritha

摘要

Large Language Models (LLMs) exhibit strong generative capabilities but remain limited in structured knowledge domains due to factual inconsistency, shallow multi-hop reasoning, and weak alignment with domain constraints. To address these limitations, this paper presents EduRAG-Compose, a unified hybrid retrieval–generation architecture that integrates hierarchical coarse-to-fine retrieval, dynamic knowledge graph traversal, cache-augmented inference, and compositional multi-step generation into a single framework for structured domain reasoning. The architecture combines clustered dense retrieval for efficient evidence selection, graph-based dependency traversal for multi-hop inference, recursive retrieval–generation cycles for compositional reasoning, and a caching layer to reduce end-to-end latency. A domain-alignment classifier and a constrained decoding strategy further support structured and interpretable output generation consistent with underlying semantic dependencies. The framework is evaluated on structured-domain benchmarks including Science Question Answering (ScienceQA), Textbook Question Answering (TQA), and interaction-derived query sets from EdNet, focusing on retrieval quality, reasoning coherence, and generative consistency. Across these datasets, EduRAG-Compose achieves 91.2 ± 1.3% factual accuracy, outperforming vanilla Retrieval-Augmented Generation (RAG) at 78.4 ± 2.1%, ColBERT-based Retrieval-Augmented Generation (ColBERT-RAG) at 82.9 ± 1.9%, and Generative Pre-trained Transformer with retrieval (GPT + RAG) at 88.5 ± 1.5%. Domain-alignment performance improves from 0.68 ± 0.05 to 0.89 ± 0.03, while structural reasoning consistency increases from 0.61 ± 0.06 to 0.87 ± 0.04, with an average response latency of 4.8 ± 0.7 s, comparable to baseline retrieval systems and substantially lower than Chain-of-Thought (CoT) prompting. All reported improvements are statistically significant based on paired Wilcoxon signed-rank tests (p < 0.01). Qualitative reasoning traces further demonstrate that explicit retrieval paths and graph traversal sequences enhance interpretability and support human-in-the-loop verification. Rather than introducing new retrieval or generation algorithms, the primary contribution lies in the systematic integration of complementary retrieval and reasoning mechanisms into a single transparent architecture for structured domains. Limitations related to dataset scope, computational overhead, and text-only processing are acknowledged, and future work will explore multimodal extensions, computational optimization for real-time deployment, and adaptive reasoning mechanisms capable of incorporating continuous feedback or evolving knowledge sources.