LeiWangetal. ASurveyonLargeLanguageModelbasedAutonomousAgents 19
Parameter Learning Parameter Learning Parameter Learning
Model
Output Model
Dataset Prompt Engineering
Prompt Engineering
Agent relevant Prompts Output
Model
Classify the text into
neutral, negative or Output Mechanism Engineering
positive. Text: I think
Input the food was okay. Neutral Trial-and- Crowd-
Sentiment: Error sourcing
Model Parameters Capability Prompts Capability Capability
The era of machine learning The era of large language model The era of agent
Fig.4 Illustrationoftransitionsinstrategiesforacquiringmodelcapabilities.
addition to building datasets based on human or (i.e., prompt engineer). In prompt engineer, one
LLMannotation,directlyusingreal-worlddatasets needstowritevaluableinformationintotheprompts
tofine-tunetheagentisalsoacommonstrategy. For toenhance themodelcapabilityor unleashexisting
example,inMIND2WEB[89],theauthorscollecta LLM capabilities. In the era of agents, the model
largeamountofreal-worlddatasetstoenhancethe capabilitycanbe acquiredbasedonthreestrategies:
agent capability in the web domain. In contrast to (1)modelfine-tuning, (2)promptengineer and(3)
priorstudies, thedatasetpresentedin thispaperen- designingproperagentevolutionmechanisms(we
compassesdiverse tasks,real-worldscenarios, and called it as mechanism engineering). Mechanism
comprehensive user interaction patterns. Specif- engineeringisabroadconceptthatinvolvesdevelop-
ically, the authors collect over 2,000 open-ended ingspecializedmodules,introducingnovelworking
tasks from 137 real-world websites spanning 31 rules, and other strategies to enhance agent capa-
domains. Using this dataset, the authors fine-tune bilities. For clearlyunderstanding suchtransitions
LLMstoenhancetheirperformanceonweb-related onthestrategyofmodelcapabilityacquisition,we
tasks, including movie discovery and ticket book- illustrate them in Figure 4. In the following, we
ing,amongothers. InSQL-PALM[90],researchers introduce prompting engineering and mechanism
fine-tune PaLM-2 based on a cross-domain large- engineeringforagentcapabilityacquisition.
scale text-to-SQL dataset called Spider. The ob-
• Prompting Engineering. Due to the strong
tained model can achieve significant performance
languagecomprehensioncapabilities,peoplecandi-
improvementontext-to-SQLtasks.
rectlyinteractwithLLMsusingnaturallanguages.
CapabilityAcquisitionwithoutFine-tuning: In Thisintroducesanovelstrategyforenhancingagent
theera oftraditionmachine learning,the modelca- capabilities,that is,onecandescribethe desiredca-
pabilityismainlyacquiredbylearningfromdatasets, pability using natural language and then use it as
wheretheknowledgeisencoded intothemodelpa- promptstoinfluenceLLMactions. Forexample,in
rameters. IntheeraofLLMs,themodelcapability CoT [45], in order to empower the agent with the
can be acquired either by training/fine-tuning the capability forcomplex taskreasoning, the authors
model parameters or designing delicate prompts presenttheintermediatereasoningstepsasfew-shot