24 Front. Comput. Sci.,2024,0(0): 1–42
haveshownstrongcapabilitiesonlanguageunder- Natural ScienceEducation: LLM-based agents
standing and employing tools such as the internet cancommunicatewithhumansfluently,oftenbeing
and databases for text processing. These capabili- utilizedtodevelopagent-basededucationaltools[115,
ties empower the agent to excel in tasks related to 117–119]. Forexample,[115]developsagent-based
documentationanddatamanagement[75,115,116]. educationsystems tofacilitatestudents learningof
In [115], the agent can efficiently query and uti- experimentaldesign,methodologies,andanalysis.
lize internet information to complete tasks such The objective of these systems is to enhance stu-
as question answering and experiment planning. dents’ criticalthinking and problem-solving skills,
ChatMOF [116] utilizes LLMs to extract impor- whilealsofosteringadeepercomprehensionofsci-
tant information from text descriptions written by entific principles. Math Agents [117] can assist
humans. Itthenformulatesaplantoapplyrelevant researchers in exploring, discovering, solving and
tools forpredicting theproperties and structuresof proving mathematical problems. Additionally, it
metal-organicframeworks. ChemCrow[75]utilizes can communicate with humans and aids them in
chemistry-relateddatabasestobothvalidatethepre- understanding and using mathematics. [118] uti-
cision of compound representations and identify lize the capabilities of CodeX [119] to automati-
potentiallydangeroussubstances. Thisfunctional- cally solve andexplain university-level mathemati-
ityenhancesthereliabilityandcomprehensiveness cal problems, which can be used as education tools
of scientific inquiries by ensuring the accuracy of toteachstudentsandresearchers. CodeHelp[120]
thedatainvolved. is an education agent for programming. It offers
manyusefulfeatures,suchassettingcourse-specific
ExperimentAssistant: LLM-basedagentshave
keywords,monitoringstudentqueries,andprovid-
the ability to independently conduct experiments,
ing feedback to the system. EduChat [86] is an
making them valuable tools for supporting scien-
LLM-based agent designed specifically for the edu-
tistsintheirresearchprojects[75,115]. Forinstance,
cationdomain. Itprovidespersonalized,equitable,
[115]introducesaninnovativeagentsystemthatuti-
andempatheticeducationalsupporttoteachers,stu-
lizesLLMsforautomatingthedesign,planning,and
dents,andparentsthroughdialogue. FreeText[121]
execution of scientific experiments. This system,
is an agent that utilizes LLMs to automatically as-
whenprovidedwiththeexperimentalobjectivesas
sess students’ responses to open-ended questions
input, accesses the Internet and retrieves relevant
andofferfeedback.
documents to gather the necessary information. It
subsequentlyutilizesPythoncodetoconductessen-
3.3 Engineering
tialcalculationsandcarryoutthefollowingexper-
iments. ChemCrow[75]incorporates17carefully LLM-based autonomous agents have shown great
developedtoolsthatarespecificallydesignedtoas- potential in assisting and enhancing engineering
sistresearchersintheirchemicalresearch. Oncethe research and applications. In this section, we re-
inputobjectivesarereceived, ChemCrow provides viewandsummarizetheapplicationsofLLM-based
valuable recommendations for experimental proce- agentsinseveralmajorengineeringdomains.
dures,whilealsoemphasizinganypotentialsafety Civil Engineering: In civil engineering, LLM-
risksassociatedwiththeproposedexperiments. based agents can be used to design and optimize