Scholarlink

The exponential growth of scientific literature, particularly on platforms such as arXiv, presents a significant challenge for researchers to efficiently locate and grasp relevant content. This paper proposes a concept-aware summarization pipeline that combines Retrieval-Augmented Generation (RAG) with semantic search and fine-tuned language models. The system employs a transformer-based concept selector to identify key concepts, enhancing retrieval relevance and summary consistency. Summary generation is performed using a lightweight, Low-Rank Adaptation (LoRA) model for its ability to efficiently fine-tune large language models with significantly fewer parameters, optimizing LLaMA 3.2 (1B) model, enabling efficient, domain-aware summarization. Semantic search is facilitated through Facebook AI Similarity Search (FAISS), supporting query-based summarization workflows. The pipeline includes a Gradio-based user interface for accessible interaction. The proposed approach addresses the limitations of generic Large Language Models (LLMs) in handling long, domain-specific documents and aims to assist researchers in accessing concise, meaningful insights from scientific papers. The system demonstrates strong performance, achieving a ROUGE-L score of 0.73 and a BERTScore F1 of 0.95.

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

LLaMA Meets RAG: Concept-Aware Summarization of Academic Papers

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Shribhakti S. Vibhuti,
  • Snehal V. Devasthale,
  • Shruti Sutar,
  • Bhoomika Marigoudar,
  • Saakshi Lokhande,
  • Uday Kulkarni

摘要

The exponential growth of scientific literature, particularly on platforms such as arXiv, presents a significant challenge for researchers to efficiently locate and grasp relevant content. This paper proposes a concept-aware summarization pipeline that combines Retrieval-Augmented Generation (RAG) with semantic search and fine-tuned language models. The system employs a transformer-based concept selector to identify key concepts, enhancing retrieval relevance and summary consistency. Summary generation is performed using a lightweight, Low-Rank Adaptation (LoRA) model for its ability to efficiently fine-tune large language models with significantly fewer parameters, optimizing LLaMA 3.2 (1B) model, enabling efficient, domain-aware summarization. Semantic search is facilitated through Facebook AI Similarity Search (FAISS), supporting query-based summarization workflows. The pipeline includes a Gradio-based user interface for accessible interaction. The proposed approach addresses the limitations of generic Large Language Models (LLMs) in handling long, domain-specific documents and aims to assist researchers in accessing concise, meaningful insights from scientific papers. The system demonstrates strong performance, achieving a ROUGE-L score of 0.73 and a BERTScore F1 of 0.95.