This project provides a framework and sandbox environment for simulating realistic user behavior using Large Language Models (LLMs) as agents. It addresses the challenge of generating high-quality user behavior data for human-centered applications, enabling researchers to study social phenomena like information cocoons and user conformity.

How It Works

RecAgent employs an LLM-based agent framework with a human-like memory mechanism, mimicking sensory, short-term, and long-term memory. This approach allows agents to efficiently manage and recall relevant information, leading to more consistent and believable behavior simulations. The system supports large-scale, parallel multi-agent simulations, integrating with OpenAI APIs for up to 1000 concurrent agents.

Quick Start & Requirements

• Install via pip install -r requirements.txt (note: faiss-cpu may require conda install faiss-cpu -c pytorch).
• Requires an OpenAI API key, configured in config/config.yaml.
• Uses MovieLens-1M for item data; user profiles and relationships are fabricated.
• Setup involves cloning the repository, installing dependencies, and configuring API keys and simulation parameters.

Highlighted Details

• Novel user simulation paradigm combining LLMs and user behavior analysis.
• Human-like memory mechanism for realistic and consistent agent behavior.
• Large-scale multi-agent simulation supporting up to 1000 agents.
• Multi-dimensional evaluation of agent and system performance.
• High extensibility for various recommendation algorithms and LLMs.
• Human-in-the-loop simulation and flexible agent intervention capabilities.

Maintenance & Community

The project has been accepted by TOIS and is nearing completion. Key contributors include Lei Wang, Jingsen Zhang, and Hao Yang.

Licensing & Compatibility

MIT License. All data and code are restricted to academic purposes only.

Limitations & Caveats

The project is primarily for academic use due to licensing restrictions. Fabricated user data may not fully represent real-world user populations.