Page 113 - Read Online

Li et al. J. Mater. Inf. 2026, 6, 10 Page 5 of 31

significantly enhances the comprehensiveness and accuracy of data extraction and enables a synergistic
mining of both the depth and breadth of literature information. By transforming unstructured text into
high-quality, structured databases, they provide interrelated foundational data that establishes a cornerstone
for materials informatics and lays a solid foundation for data-driven materials discovery.

However, fully transforming this potential into a mature paradigm still faces two key challenges. Firstly, on
the technical front, although agents can already process multimodal information such as text, charts, and
microscopic images in parallel, achieving unified semantic representation and cross-modal deep reasoning of
these heterogeneous data remains a core challenge. Semantic gaps and structural differences exist between
different modalities, and current cross-modal fusion mechanisms are still immature. These often lead to
information loss or erroneous associations during information alignment and joint reasoning, thereby
limiting the completeness and accuracy of overall knowledge extraction. Secondly, at the application
closed-loop level, systematic solutions are still lacking for seamlessly integrating the extracted massive
structured data into the complete workflow, from material design and performance prediction to
experimental validation. Therefore, in the future, it is necessary to further explore cross-modal pre-training,
knowledge-enhanced multimodal alignment, and intelligent decision-making systems embedded in the
material research and development workflow, so as to truly unlock the hidden knowledge value in scientific
literature and promote the transformation of material research from data-rich to knowledge-intelligent.

Knowledge question answering
Based on structured data, there is a growing need to efficiently and accurately integrate knowledge scattered
in multi-source databases to provide professional answers to questions in material science. The approach of
using general-purpose LLMs alone often encounters problems such as outdated knowledge, limited domain
expertise, and insufficient interpretability. AI agents offer a new approach to building powerful materials
knowledge question answering systems. By integrating techniques such as retrieval-augmented generation
(RAG), tool invocation, and dynamic knowledge base update, agents significantly enhance the ability to
acquire and reason with professional knowledge. They can retrieve, reason, and generate precise answers
from vast amounts of professional information, thereby ensuring responses that are both timely and
scientifically reliable.

Specifically, agents employ RAG to dynamically acquire the latest information from external knowledge
bases, effectively compensating for potential obsolescence or insufficiency in the LLM’s internal knowledge.
For instance, Mixture-of-Workflows for Self-Corrective Retrieval-Augmented Generation (CRAG-MoW)
achieved performance comparable to GPT-4o (“o” for “omni”) in question answering and retrieval of
multimodal materials data such as molecules, polymers, chemical reactions, and nuclear magnetic resonance
(NMR) spectra by adopting a multi-agent workflow and self-correcting RAG strategy, highlighting the
advantages of structured retrieval and MAS . Similarly, the Collaborative framework of LLM Agents for
[29]
Drug Discovery (CLADD) generates high-quality responses without domain fine-tuning by collaborating
with multiple agents and dynamically retrieving biomedical knowledge bases to contextualize query
molecules . MechGPT combines knowledge graphs with RAG to enhance the reasoning capability and
[30]
interpretability of agents in question answering in interdisciplinary domains . These studies indicate that
[31]
multi-agent coordination coupled with RAG can effectively improve the quality of complex materials
knowledge question answering. Under this collaborative framework, the key to achieving efficient reasoning
and accurate answers across modal and cross-domain data lies in how to coordinate the division of tasks
among agents. In practice, hierarchical decision-making and role division mechanisms are often adopted.
For instance, the “coordinator agent” is responsible for task parsing and allocation, while “specialist agents”
handle information of different modalities, such as text, images, or structured data, independently. These
agents leverage RAG to achieve complementary knowledge and cross-validation. This division not only