LifeSim - A long-term user life simulation framework launched by Fudan University and Shanghai Innovation Academy

What is LifeSim?

LifeSim is the first long-term user life simulation framework launched by Fudan University and Shanghai Institute of Innovation and Entrepreneurship to evaluate personalized AI assistants. Based on the BDI cognitive model, the framework simultaneously models the user's internal cognition (beliefs, desires, intentions) and the external environment (time, location, weather) to generate coherent life trajectories and multi-turn interactions. Through the LifeSim-Eval benchmark across 1200 scenarios and 8 life domains, LifeSim evaluates the model's ability to handle explicit and implicit intentions and model long-term preferences, addressing the problem of existing evaluations being disconnected from real-world scenarios.

LifeSim's main functions

• Long-distance life trajectory simulationIt generates a coherent sequence of user life events based on real travel data, covering external environmental constraints such as time, location, and weather.
• Multi-turn interaction behavior simulationIt simulates natural conversations between users and AI assistants, and supports memory conflict detection, emotion reasoning, and dynamic behavior selection.
• Personalized Ability AssessmentThe LifeSim-Eval benchmark model identifies explicit/implicit intentions and the ability to reconstruct and align long-term preferences.
• Privacy and security data synthesisIt supports the generation of millions of diverse user profiles, providing high-quality synthetic data for personalized assistant training.

LifeSim's technical principles

• Belief EngineIntegrate long-term user profiles with short-term situational cognition (physical/psychological/environmental state).
• Wish EngineRetrieve candidate intents from the demand database and reorder them based on user beliefs and the external environment.
• Event EngineThe probability of event triggering is controlled based on logical functions to ensure that real-life events conform to real-world spatiotemporal constraints.
• Behavior EngineUser responses are generated through three stages: memory perception (detecting historical conflicts), emotion reasoning (GoEmotions classification), and action selection.

How to use LifeSim

• Online Demo Experience:
  • Visit the official websiteYou can access the visual interface directly at http://fudan-disc.com/lifesim/.
  • Preset DemonstrationSelect a user built into the system, click on any point in the life on the map timeline, view the scene at that time, and have a conversation with the simulated user.
  • Real-time generationCustomize your profile with age, occupation, personality traits, etc., and watch and interact with real-time life events generated by the BDI engine.
• Local Deployment Evaluation:
  • Environmental preparationInstall Python dependencies and prepare user profiles and life event data.
  • Model Configuration: Connect user simulation models (such as Qwen3-32B) and the assistant model under test (supports local vLLM or APIs such as OpenAI/DeepSeek).
  • Run simulationIf you choose a single-scene (independent dialogue) or long-term mode (multi-turn interaction with history memory), the system will automatically generate a dialogue log.
  • Automatic ratingThe LLM-as-Judge algorithm is used to automatically score seven dimensions, including intent recognition, implicit need fulfillment, and profile alignment.

Key information and usage requirements for LifeSim

• positionThe first long-duration user life simulator used to evaluate the capabilities of personalized AI assistants in real-world dynamic scenarios.
• core technologyBased on the BDI cognitive model (belief-desire-intention), user behavior is generated by integrating external environment (time/location/weather) and internal cognitive state (personality/preference/memory).
• Data scaleA pool of millions of user profiles, 3,374 real trajectory data, and 1,200 evaluation scenarios (covering 8 major life areas).
• Evaluation FocusDistinguish between explicit intent (direct need) and implicit intent (requires inference based on user profile and context); support long-range dialogue (up to 16K tokens of historical context).
• AffiliationDeveloped jointly by Fudan University's Data Science Department and Shanghai Innovation Academy.

LifeSim's core advantages

• Filling the gap in real-world scenariosBreaking through the limitations of existing static and short-context evaluations, it achieves for the first time long-distance life trajectory simulation across days/weeks, making AI assistant evaluations truly close to "Jarvis"-style continuous service scenarios.
• BDI Cognitive Architecture Deep ModelingUnlike simple role-playing, this system is based on a belief-desire-intention psychological model, which simulates how users "think" (cognitive state) and generates a reasonable behavioral chain with consistent personality.
• Double test of explicit and implicit intentionsThe implicit intent (which needs to be inferred from weather, location, and historical preferences) exposes the 20+ point performance gap in the current LLM in long-term user understanding.
• Integration of physical and cognitive environmentsIt uniquely combines real geographical trajectories (3,374), physical constraints such as time and weather, with cognitive mechanisms such as dynamic psychological emotions and memory forgetting to generate events with realistic rationality.

LifeSim project address

• GitHub repository: https://github.com/dfy37/lifesim
• arXiv technical paper: https://arxiv.org/pdf/2603.12152
• Experience the demo onlinehttp://fudan-disc.com/lifesim/

LifeSim's Competitive Product Comparison

LifeSim Application Scenarios

• AI Assistant Capability Evaluation and BenchmarkingIt provides a standardized "long-range personalized" testing environment for models such as GPT-4o, Claude, and DeepSeek, accurately identifying the limitations of the models' capabilities in areas such as implicit intent understanding, long-term memory retention, and user profile alignment.
• Synthetic data generation:Based on large-scale, diverse, long-term interactive dialogue data generated from millions of simulated users, it solves the problems of scarcity and privacy sensitivity of real user data, and can be used to fine-tune personalized assistants or reinforcement learning training.
• Intelligent Customer Service and Companion AI Pre-trainingSimulate extreme or rare scenarios in a virtual environment (such as a user seeking help in a state of anxiety for several consecutive days) to test the emotional support capabilities and long-term contextual consistency of the customer service system, avoiding the risks of testing with real users after launch.
• Human-Computer Interaction (HCI) Academic ResearchIt provides a controllable experimental platform for cognitive science and social psychology to study how different personality traits (such as the Big Five personality traits) affect users' acceptance of AI assistants and the process of building trust.
• Personalized recommendation algorithm verificationIn eight major areas of life, including diet, fitness, and parenting, we will verify whether the recommendation system can make dynamic adjustments based on users' long-term preferences and real-time contexts (such as rainy days + fitness habits).