Towards Faithful Simulation of Human Shopping Behavior
RecVerse is built to replay shopping sessions more like real users by combining hierarchical memory with session-level reinforcement learning.
The paper says existing LLM and VLM shopping simulators struggle with long browsing histories and step-by-step training that misses whole-session behavior. RecVerse uses working, episodic, and preference memory, with memory updates treated as actions. It is trained against trajectory-level signals covering both action patterns and shopping intent. The authors also release USB, an interactive e-commerce GUI trajectory dataset, and report stronger fidelity and intent consistency than baselines. HF Daily Papers' note
The paper says existing LLM and VLM shopping simulators struggle with long browsing histories and step-by-step training that misses whole-session behavior. RecVerse uses working, episodic, and preference memory, with memory updates treated as actions. It is trained against trajectory-level signals covering both action patterns and shopping intent. The authors also release USB, an interactive e-commerce GUI trajectory dataset, and report stronger fidelity and intent consistency than baselines. HF Daily Papers' note
score 4