Collaborative Optimization Approach for Workflow Agents in User Behavior Modeling

User behavior modeling is increasingly critical for personalized services and decision-making systems, yet integrating diverse user and product features into a coherent review generation process remains challenging. Thus we propose a novel collaborative optimization approach for workflow agents in user behavior modeling that integrates user and product feature extraction with review and rating generation. Firstly, to solve the difficulty in determining the optimal structure of agent workflows, we employ Monte Carlo Tree Search (MCTS) to optimize the workflow architecture, establishing a high-performance baseline. Meanwhile, to tackle the challenges in generating and optimizing single-agent prompts and demonstrations, we implement a heuristic optimization strategy for joint automated tuning of system prompts and demo cases. Furthermore, through comprehensive analysis of data distributions, we construct a dynamic routing mechanism for the agent workflow, achieving enhanced performance across diverse scenario-specific datasets. We validate the effectiveness of our methods on three real-world datasets, demonstrating significant performance improvements across all proposed techniques. This approach secured second place overall (and first in star-rating prediction) in the user modeling track of the AgentSociety Challenge @ WWW 2025.