From AI Scientists to Social Science: Tsinghua Unveils AgentSociety² Silicon‑Based Social Lab

AgentSociety² is a next‑generation platform that integrates large‑language‑model (LLM) agents, realistic social environments, and a high‑performance simulation engine to create a "silicon‑based" society that can be used as an executable laboratory for social‑science research.

The system extends the earlier AgentSociety‑1, which demonstrated how AI agents could form a simulated society, by adding a fully integrated research environment. It allows researchers to observe how collective dynamics emerge from individual interactions within a controllable, reproducible computational setting.

Two core roles are defined: AI Social Scientists , which assist researchers in literature review, hypothesis generation, experiment design, and result analysis; and Silicon Participants , which act as the simulated social actors whose behavior and data are generated within the environment.

AgentSociety² provides an end‑to‑end workflow that links every stage of a social‑science project—topic definition, literature search, hypothesis formulation, experimental configuration, simulation execution, result interpretation, and paper writing—into a single auditable pipeline. Human researchers remain in the loop at critical decision points, such as revising hypotheses, setting parameters, and interpreting outcomes.

Unlike traditional AI‑Scientist systems that rely on static prompts, AgentSociety² adopts a skill‑based architecture. Observation, cognition, planning, memory, and domain‑specific decision rules are modularized as reusable skills that agents load on demand. Each agent maintains an independent workspace to store its profile, state, memory, logs, and checkpoints, supporting long‑duration, traceable simulations.

The platform also modularizes social environments (public‑goods games, prisoner's dilemma, trust games, psychological experiments, social‑media spaces, event‑driven scenarios, economic and mobility models) as "agentic environments". Researchers can compose these modules to construct custom experiments, and the CodeGenRouter translates natural‑language intents into validated environment operations, eliminating the need to write low‑level simulation code.

To demonstrate versatility, the team designed seven experiment categories spanning micro‑behavioral studies, meso‑level network dynamics, and macro‑level urban and disaster scenarios. Examples include modeling opinion polarization in social media, investigating cooperation decay in public‑goods games, and simulating population movement during disaster response.

AgentSociety² thus provides a new infrastructure for computational social science, enabling the construction of synthetic populations, configurable environments, and systematic interventions that can be run, compared, and reproduced. It emphasizes a human‑in‑the‑loop paradigm where AI expands the mechanism space while researchers retain judgment and control.

Future applications are envisioned in platform governance, public policy, city management, disaster response, collective decision‑making, social psychology, and AI safety, offering a controllable experimental sandbox that complements, rather than replaces, real‑world observations.