Scholarlink

Extracting multi-modal information from scientific documents can help answer questions in scientific research. This paper presents a multi-modal agent framework that combines structured summarization with multi-vector retrieval to enable cross-modal information extraction and reasoning for materials science research. Experimental results on the SciAssess benchmark demonstrate that the proposed approach outperforms strong LLM baselines, across multiple scientific tasks, such as alloy composition extraction, chart-based question answering, and procedural information understanding. This validates the effectiveness and versatility of the method, laying a foundation for the development of a scientific assistant.

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

A Retrieval-Augmented Multimodal Framework for Scientific Reasoning in Materials Science

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Ziyi Luo,
  • Jian Xu,
  • Qingbo Yan,
  • Cheng-Lin Liu

摘要

Extracting multi-modal information from scientific documents can help answer questions in scientific research. This paper presents a multi-modal agent framework that combines structured summarization with multi-vector retrieval to enable cross-modal information extraction and reasoning for materials science research. Experimental results on the SciAssess benchmark demonstrate that the proposed approach outperforms strong LLM baselines, across multiple scientific tasks, such as alloy composition extraction, chart-based question answering, and procedural information understanding. This validates the effectiveness and versatility of the method, laying a foundation for the development of a scientific assistant.