王磊_RecAgent-20_watermark.pdf

1、When Large Language Model based Agents meet User Behavior SimulationRenmin University of China Lei WangDifferent Study Paradigms in AI Real-world Datasets A large amount of public available datasets The datasets can be easily collectedSimulation EnvironmentsThe data generation mechanisms are known o

2、r can be accurately predictedWhere is the Position of User Behavior AnalysisUser Behavior AnalysisSimulation EnvironmentsReal-world datasetsRecommendation SystemSocial NetworkUser Behavior Tracking A large amount of public available datasets The datasets can be easily collectedThe data generation me

3、chanisms are known or can be accurately predictedUser PrivacyCommercial Confidentiality Ethical ProblemIntricate Human Cognitive ProcessComplex EnvironmentsComplex Influential FactorsSimulation based User Behavior AnalysisSimplified User Decision ProcessesSimplified EnvironmentsRely on Real-world Da

4、tasets Clicking Behaviors Browsing Behaviors Purchasing Behaviors Watching Behaviors Inner Product Multilayer Perceptron Deep Neural Network Real-world DatasetsWatching MoviesOpinion SharingSimple environments Simple functions Sparse datasetsUnreliable SimulationDifferent Environments May Mutually I

5、nfluence Each OtherSimulatorTrainingUser BehaviorsZero-Shot Simulation is ImportantA Novel Paradigm for User Behavior SimulationThe Fast Growing of Large Language Models Human-level Intelligence Simplified User Decision ProcessesSimplified EnvironmentsRely on Real-world Datasets Surprisingly Strong

6、Generalization CapabilityHuman-level Intelligence Generalization CapabilityUnified NLP interfaceZero-shot InferenceHuman-like decisionsZero-shot InferenceAgent-level DesignAgent=LLM+Profiling Module+Memory Module+Action ModuleProfiling Module Agent-level DesignMemory Module Agent-level DesignRichard