RecVerse Agent Simulates Human Shopping Behavior

Researchers from the Renmin University of China, UCAS, NUS, and the Alibaba Group have developed RecVerse, a GUI-grounded simulation agent designed to mimic how real people browse and shop online. Unlike traditional simulators that struggle with memory overload or unrealistic, repetitive actions during long browsing sessions, RecVerse interacts directly with user interfaces using screenshots. This allows the agent to generate highly realistic, multi-turn shopping trajectories that can be used to train and evaluate recommendation algorithms without risking real-world user traffic.

To overcome the limitations of previous models, RecVerse utilizes a cognitively inspired hierarchical memory structure. This system is divided into three parts: a first-in, first-out working memory for recent screenshots and immediate mindset; an episodic memory that logs session-level textual event traces; and a preference memory that distills long-term user intent. The agent treats memory updates as actions, learning what to remember through reinforcement learning. Additionally, its trajectory-aligned reinforcement learning framework utilizes macro rewards to align action types, micro rewards based on a three-tier category tree for intent signals, and format rewards to maintain valid outputs.

Alongside the agent, the researchers released the User Simulation Benchmark (USB), which they describe as the first interactive GUI benchmark supporting multi-turn agentic reinforcement learning. The USB dataset contains 5,274 trajectories, 69,842 actions, 5,222 users, and 90,095 items classified across 41, 517, and 2,256 categories. In evaluations against the leading GUI baseline, STA, RecVerse demonstrated substantial improvements, boosting F1 scores by 68 percent, hit rate by 77 percent, and human-curated oracle metrics by 41 percent. Furthermore, in head-to-head human evaluations, RecVerse was judged superior to STA in 92 percent of test cases.

For e-commerce developers and machine learning practitioners, RecVerse offers a powerful tool for offline evaluation and counterfactual analysis. Instead of running expensive and potentially disruptive online A/B tests on live traffic, engineers can now deploy this high-fidelity simulator to safely test new recommendation models and reinforcement learning agents in a realistic, interactive environment.