From Blueprint to Reality: Modeling and Applying Putnam’s Social Capital Theory with LLM-based Multi-agent Simulations

Putnam's Social Capital Theory is a foundational framework explaining how communities overcome collective-action dilemmas through social networks, trust, and norms. However, traditional empirical research methods struggle with high resource demands, strict control limits, and replication barriers, while existing LLM-based social simulations remain behavior-driven rather than theory-aligned. To address these gaps, the authors introduce SocaSim, an LLM-based multi-agent simulation framework that bridges social science theory with computer science simulation by explicitly modeling the co-evolution of social networks, trust dynamics, and norm propagation.

SocaSim incorporates specialized structural and cognitive modules to give LLM agents sociologically plausible behaviors. The Social Structure Trait (SST) module initializes agents with demographic and socioeconomic status (SES) attributes sampled from real-world survey data, dictating their social capital preferences. The Belief-Desire-Intention (BDI) framework drives round-by-round decision-making by combining structural features with real-time situational cues. Furthermore, the Social Cognitive Memory (SCM) module enables adaptive learning, multi-layer trust initialization, and norm formation and decay. Agents operate in a two-phase round-based workflow consisting of a Proposal phase and an Execution phase.

The authors evaluate SocaSim through two progressive tasks: theory modeling and real-world application. In the modeling experiments, 25-round simulations demonstrate that network density co-evolves with cooperation rates, trust grows steadily across all socioeconomic groups, and reciprocity norms sustain long-term cooperation pairs. When benchmarked against 20 real human participants across eight scenario-based decisions, the LLM-based multi-agent framework exhibits strong group-level alignment with a Pearson correlation coefficient of 0.974. In the application task concerning smart elderly care, counterfactual interventions reveal that raising initial trust among low-SES agents significantly boosts technology adoption and reduces decision contradictions.