Page 114 - Read Online

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

enhances the parallel efficiency of the system in processing complex queries but also improves the accuracy
and robustness of answers through the fusion of multi-source information. However, collaboration between
multiple agents is not inherently efficient. It needs to rely on well-defined communication protocols and
dynamic task scheduling strategies to avoid information redundancy or decision conflicts, thereby fully
leveraging the ability to integrate cross-modal and cross-domain data.

To enhance the professionalism of LLMs and extend their capability boundaries, domain tools have been
deeply integrated into AI agents, enabling them to invoke professional tools for precise calculation and
validation. For example, Chemistry Agent Connecting Tool Usage to Science (CACTUS), by integrating
tools such as RDKit, significantly improves accuracy on question answering benchmarks and enables more
reliable molecular property prediction and similarity search . CheMatAgent incorporates 137 chemical
[32]
tools and the ChemToolBench dataset, and employs a hierarchical evolutionary Monte Carlo Tree Search
framework to optimize tool planning and execution, achieving an accuracy in materials question answering
that surpasses GPT-4o [Figure 2]. GVIM utilizes an LLM fine-tuned on chemical domain data and
[33]
integrates professional tools for molecular visualization and literature retrieval, combined with knowledge
accumulation and group collaboration mechanisms, to enhance the professionalism and interactivity quality
of question answering . These works highlight the importance of tool invocation for tasks requiring precise
[34]
calculation. However, the use of tools may not consistently yield benefits. Its effectiveness highly depends on
the type of task. A comparative study by ChemToolAgent reveals that for highly specialized tasks such as
synthesis prediction, equipping agents with dedicated tools can achieve significant improvement. In contrast,
for general chemical knowledge questions (e.g., exam questions), the agent’s ability to correctly reason about
chemical knowledge is more important. In such cases, tools do not always provide assistance and may even
degrade performance due to superfluous integration . Therefore, when invoking specialized tools, it is
[35]
necessary to carefully define their usage boundaries to prevent negative effects. A feasible principle is
on-demand invocation and moderate integration. The system should first identify and classify the intent of
the query. If it is a conceptual or fundamental theoretical question, the system should prioritize relying on
the enhanced knowledge base and reasoning modules. Specialized tools should only be triggered when
dealing with complex calculations, structural validation, or real-time data retrieval. Additionally, a
confidence assessment mechanism for tool outputs needs to be established. If significant discrepancies arise
between tool results and semantic reasoning, secondary verification or selective adoption should be applied.
Through such conditional tool invocation strategies, the system not only can leverage the advantages of
specialized tools in precision tasks, but also effectively avoids performance degradation and resource waste
caused by over-reliance or improper invocation.

Building upon this progress, the integration of MAS with dynamic knowledge bases and self-evolving
knowledge graphs enables the development of more adaptive and continuously learning question answering
systems, offering deeper interpretability and long-term sustainability. For instance, ChemAgent constructs a
dynamic and self-updating memory library by decomposing chemical tasks into subtasks and storing them as
structured memories. This enables the agent to retrieve and refine relevant information when addressing new
problems, thereby significantly enhancing its performance in complex chemical reasoning tasks, achieving a
performance improvement of up to 46% on the SciBench chemical reasoning dataset . TopoMAS
[36]
(Topological materials Multi-Agent System) integrates the entire materials discovery workflow, from query
and data retrieval to first-principles validation, into a dynamic knowledge graph, achieving continuous
knowledge refinement and iteration. This provides support for answering complex questions involving
material design and verification . Such approaches transform AI agents from static question answering
[37]
systems into dynamic systems capable of accumulating and evolving knowledge.