SimWorld: Human and Environment Simulation for Real Intelligence

Simulated Worlds, Real Intelligence

Open Call

Call for Program Committee Members

We invite researchers and practitioners in AI agents, user simulation, HCI, recommender systems, and evaluation to join the workshop Program Committee.

About The Workshop

Many applications depend on understanding human behavior, from human–AI collaboration and recommender systems to market research and usability testing. Human interaction data supports both system evaluation and model training, providing experience and feedback for reinforcement learning and recursive self-improvement. Yet collecting this data is slow, costly, and difficult to scale. LLM-based digital humans offer a scalable proxy: they can express preferences, pursue goals, navigate interfaces, and interact with one another to simulate collective behavior. Simulated environments, spanning websites, mobile apps, tools, and robotics, provide stable and controllable settings for these interactions. However, both face gaps in realism: digital humans may be overly cooperative, lack behavioral diversity, or misrepresent populations, while simulated environments may miss real-world complexity and fail to support generalization.

SimWorld brings together researchers and practitioners to study human and environment simulation as interrelated challenges. Realistic interactions require both diverse human behavior grounded in real data and faithful environment dynamics. The workshop connects communities across machine learning, NLP, HCI, recommender systems, multi-agent systems, robotics, and behavioral science to advance human behavior modeling, environment construction, joint evaluation, behavioral alignment, reinforcement learning, and agent self-improvement. Together, we aim to establish when simulated interactions support valid downstream decisions, understand where simulation fails, and develop responsible approaches to scalable, reproducible training, evaluation, and product design without treating synthetic proxies as replacements for human participants.

Themes

Human Behavior Modeling

How can we build digital human foundation models that faithfully reflect human goals, preferences, constraints, uncertainty, frustration, and behavioral variation across domains? How can multi-agent simulation capture interactions and collective behavior?

Evaluation and Benchmarking

What metrics, datasets, benchmarks, and protocols are needed to measure simulator quality, such as realism, diversity, controllability, robustness, and fairness?

Environment Simulation

How can we construct realistic yet controllable sandboxes for digital and robotic environments?

Reinforcement Learning and Agent Self-Improvement

How can simulated humans and environments support reinforcement learning, adaptive curricula, and recursive self-improvement?

Behavioral Alignment

How can clickstreams, interaction logs, surveys, user studies, and other behavioral data be used to improve the fidelity and validity of user simulators?