1 Introduction↩︎

When an accident occurs, large anonymous crowds often freeze, each waiting for someone else to move first. In close‑knit groups, help starts immediately because people are connected, trust one another, and expect to assist. This contrast shows how social network, trust, and norms shape a group’s ability to respond to common problems. Social Capital Theory, developed by Robert D. Putnam putnam1993?, putnam2000?, captures these dynamics and explains why some communities overcome collective‑action dilemmas while others do not.

In social science, research on Putnam’s Social Capital Theory largely relies on conventional empirical methods that yield valuable insights yet have notable limits (Figure 1 (a)). Quantitative methods such as large-scale surveys [1] and Structural Equation Modeling (SEM) hoa2021?, sumi2025? require substantial time and resources, and face practical constraints on control and replication. In parallel, simulation-based approaches such as Agent-Based Modeling (ABM) shenk2019? allow controlled tests of assumptions, yet traditional rule-based agents rely on predefined and simple decision rules and struggle to capture the complexity of human reasoning, emotion, and context-dependent behavior.

Recently, advances in large language models (LLMs) have enabled the simulation of human‑like intelligence friha2024?, wang2025?, prompting growing interest in LLM‑based agents for social simulation piao2025?, yang2025?. However, existing approaches are largely behavior‑driven rather than theory‑driven (Figure 1 (b)): they typically test whether agents exhibit plausible or human‑like behavior on specific tasks [2]–[4], but rarely construct environments aligned with core theoretical propositions. Furthermore, current frameworks seldom provide reproducible, controllable settings for such theory‑driven analysis, for example, by modeling links between network connectivity and collective outcomes or stepwise trust growth. Consequently, they still make it difficult to trace how trust accumulates, how norms are internalized, or how individual–collective tensions arise, leading to limited procedural interpretability.

To bridge this gap, we introduce SocaSim, a dynamic and theory‑grounded LLM-based multi-agent simulation framework designed to model and apply Putnam’s Social Capital Theory (Figure 1 (c)). Specifically, we construct an agent society that integrates mechanisms of dynamic social network evolution, trust adaptation, and norm diffusion, in which agents engage in repeated collective-action simulations. Each agent is endowed with basic demographic attributes and social capital endowments, and is capable of reasoning and decision-making by incorporating social relationships, interaction history, and situational context. Our agent society runs in rounds with two phases: Proposal, where agents make proposals; and Execution, where they simultaneously decide whether to act based on trust, norms, and network ties.

Based on this framework, we conduct two progressive tasks. In the modeling task, we run multi-round collective-action experiments and human-alignment validation. The simulation reproduces the macro-level patterns predicted by Putnam’s theory, and a comparison with real older adults shows strong human-agent alignment in group-level decisions (Pearson r = 0.974). Subsequently, in the applying task, we move the theory from blueprint to reality by applying its three core dimensions to adaptation challenges in smart elderly care, where technology adoption hinges on social networks, trust, and norms. In counterfactual simulations that raise low-SES agents’ initial trust, the adoption rate increases by 15.4% and decision contradictions decrease by 25.5%, suggesting that trust can serve as a key causal lever. Unlike traditional methods that are often limited to correlational conclusions, SocaSim uses LLM agents to reveal micro-level causal chains of how trust accumulates, norms are internalized, and decision contradictions emerge, offering process‑level interpretability and practical guidance for policy design.

Our contributions can be summarized as follows:

\(\bullet\) We propose SocaSim, a new interdisciplinary research paradigm using LLM‑based agents to model Putnam’s Social Capital Theory.

\(\bullet\) We develop a multi-agent framework that models the dynamics of social capital, including social network change, trust dynamics, and norms diffusion.

\(\bullet\) We conduct several experiments with macro‑level replication, human alignment, and counterfactual intervention, validating SocaSim’s explanatory power for theory and real-world issues.

2 Preliminary↩︎

To clarify the theoretical basis, we briefly introduce Putnam’s Social Capital Theory (see Appendix 10 for more examples). This theory is commonly used to explain how groups deal with collective-action dilemmas, namely situations where a shared problem is hard to resolve because coordination is costly, trust is insufficient, or long-term participation is hard to sustain. Putnam argues that social capital helps alleviate such dilemmas and improves a group’s ability to respond to common challenges.

Putnam distinguishes between two primary forms of social capital. Bonding social capital develops within close‑knit groups and strengthens internal support and solidarity. Bridging social capital connects different or loosely linked groups and facilitates information exchange and access to external resources. Conceptually, social capital can be understood through three interrelated dimensions: social network, which provides channels for social interaction and coordination; trust, which reduces uncertainty and lowers the perceived cost of collective response; and norms, which encourage reciprocity, mutual support, and sustained participation within the group. These dimensions correspond directly to the core components of our simulation: social network shapes interaction opportunities, trust guides expectations about others’ actions, and norms regulate behavior over time.

Based on these two primary forms and three core dimensions, we leverage LLM agents to model and apply Putnam’s Social Capital Theory.

3 Methods↩︎

3.1 Overview of SocaSim↩︎

Since Putnam’s Social Capital Theory involves complex social constructs (social network, trust, norms), prior general multi-agent frameworks cannot capture their dynamic coupling. We therefore propose SocaSim, an LLM‑based multi‑agent framework to model, dynamically couple, and validate the theory’s core elements. Specifically, we construct a social simulation system consisting of \(N\) agents, denoted as \(A = \{a_1, \dots, a_N\}\), each equipped with demographic and socioeconomic attributes and initial social‑capital endowments, situated in a controlled environment \(E\). The simulation runs for \(T\) discrete rounds. In each round \(t\) (\(t = 1, \dots, T\)), agents initiate or respond to social behaviors based on their network connections, trust dispositions, and perceived norms. These interactions update three interrelated core dimensions:

\(\bullet\) Social network \(\boldsymbol{G_{t}}\): Connections dynamically form, strengthen, or weaken over time based on interaction frequency and quality.

\(\bullet\) Trust \(\boldsymbol{T_{i}(t)}\): Each agent \(a_{i}\) evolves its level of specific trust toward specific others and general trust toward the broader group through interactions.

\(\bullet\) Norms \(\boldsymbol{N_{i}(t)}\): Each agent’s reciprocity norms comprise specific reciprocity for direct resource exchange and general reciprocity for mutual cooperation without expecting immediate return.

As shown in Figure 2, these abstract dimensions are concretely implemented by the agent architecture (i.e., SST, BDI, and SCM) and the LLM‑driven updating mechanisms detailed in §3.2.

3.2 LLM-based Multi-agent Framework↩︎

3.2.1 Social Structure Trait (SST)↩︎

Putnam’s Social Capital Theory holds that socioeconomic position determines access to networks and resources, yet most multi-agent frameworks lack ways to model this. To address the limitation, we design a module that generates agents’ structural social features through three key components:

Basic Demographic Attributes. Each agent is assigned demographic traits, including age, gender, education, income, health status, and occupation. These attributes are sampled from the 2023 Chinese General Social Survey (CGSS) data¹ elderly subsample, ensuring that the simulated population reflects real-world socioeconomic stratification.

Socioeconomic Status (SES). A weighted composite of income, education, and occupation, classifying agents into low, mid, and high SES categories.

Social Capital Type Preference. Low‑SES agents tend to rely on bonding capital, high‑SES agents are more adept at leveraging bridging capital, while mid‑SES agents exhibit a balanced preference.

All these attributes are encoded into the agent’s natural language prompt, giving each agent a rich identity and ensuring behavior arises from sociologically plausible profiles rather than static rules.

3.2.2 Belief-Desire-Intention (BDI)↩︎

To capture the real-time decision-making of agents based on structural features from the SST module, we adopt the BDI framework bratman1999?. Specifically, we formalize the agent’s context-sensitive cognitive decision process, with \[\text{Act} \gets \mathrm{BDI}(f, B, D, I, S, M),\] where \(\text{Act}\) denotes the final action, \(f\) is implemented with an LLM, and \(B, D, I, S, M\) denote the Belief, Desire, Intention, SST structures, and SCM memory, respectively. Belief combines structural cues such as SES and education with situational cues from the current interaction round. Desire represents motivations such as technological mastery and family recognition that are modulated by the agent’s structural traits. Intention integrates belief, desire, and memory to produce concrete behavioral commitments; for technology adoption, agents evaluate their SES, prior experience, social influence, and family recommendation.

3.2.3 Social Cognitive Memory (SCM)↩︎

As agents learn through repeated interactions, they adapt their cognitive states by integrating new experiences with stored memories. Therefore, we design the SCM module to model this adaptive learning process. Specifically, we define its cognitive update process as \[(C_{\text{new}},\; M_{\text{new}}) \gets \operatorname{SCM}\bigl(\,f,\; Act,\; O,\; C_{\text{old}},\; M_{\text{old}}\,\bigr),\] where \(C_{\text{new}}\) denotes the updated cognitive state, \(M_{\text{new}}\) represents the updated memory structure, \(Act\) is the action previously executed by the agent (i.e., the output of the BDI module), \(O\) is the observed outcome, \(C_{\text{old}}\) and \(M_{\text{old}}\) are the prior states, and \(f\) is the same LLM used in the BDI module.

Multi-layer Trust Initialization. Agents maintain specific and general trust. Specific trust is initialized based on social distance. Higher for the same group, moderate for adjacent groups, and lower for distant groups, with small random noise. During the simulation, trust levels are updated according to interaction outcomes. Positive interactions such as promise keeping or providing help increase trust, while negative interactions such as default or deception decrease trust, ensuring that trust values always evolve within a meaningful range.

Norm Formation and Decay. Normative cognition comprises general and specific dimensions. Initial norm levels correlate with SES. To capture the socially reinforced nature of norm maintenance, our system checks each agent’s reciprocity behavior every round. If no reciprocity is observed over ten consecutive rounds, norm decay is triggered.

Memory and Reflection. We divide the agent’s memory into two distinct layers, interaction-history records specific interaction events, participants, and outcomes; cognitive-state stores the evolving trajectories of trust and norms. After each round, an LLM-driven reflection process analyzes these memories each round, adjusting the agent’s state and strategies for continuous adaptation.

3.3 Multi-agent Simulation Environment↩︎

3.3.1 Social Network and Capital Evolution↩︎

We design the social network to co-evolve interactively with the accumulation of social capital, forming a synergistic feedback loop during the simulation. Initially, ties are biased toward SES‑similar agents. Network density, computed as \(Density = \frac{2|E|}{|V|(|V|-1)},\) where \(|E|\) is the number of social ties and \(|V|\) is the number of agents, measures the overall connectedness of the society. During the simulation, agents accumulate bonding social capital from strong‑tie interactions and bridging social capital from weak‑tie or cross‑group interactions. The update mechanism of social capital takes into account an agent’s capital type preference. When the type of interaction an agent engages in matches its own social capital preference, the accumulation rate is significantly accelerated; otherwise, it accumulates at a baseline rate.

3.3.2 Two-phase Round-based Decision↩︎

We extend turn‑based simulations with a two‑stage Proposal–Execution process, capturing the social progression of collective action from initiation to fulfillment. Each round follows this sequence.

Proposal phase. Randomly selected initiators propose collective actions or technology adoptions. When an agent \(a_i\) proposes, each neighbor \(a_j\) responds using LLM-based reasoning that incorporates \(a_j\)’s social attributes, especially its specific trust toward \(a_i\) and its perception of general norms. This captures how social attributes shape an agent’s initial inclination toward collective decisions.

Execution phase. After accepting a proposal, each participant \(a_i\) privately decides whether to honor the commitment, again using LLM-based reasoning that considers its current trust, norms, and social context. The collective action succeeds if the proportion of committing participants exceeds a threshold \(\tau\) (default \(\tau = 0.5\)), with \[\text{Success} = \mathbb{1}\left[\frac{|\text{Committers}|}{|\text{Participants}|} > \tau \right].\]

4 Experimental Setup↩︎

4.1 Task Formulation↩︎

To evaluate the capabilities of SocaSim for Putnam’s Social Capital Theory, we conduct two progressive tasks that examine how agents embody the dynamic coupling of social capital in interactions:

\(\bullet\) Modeling. Can LLM agents reproduce the core dimensions of the theory and maintain consistency in group‑level decision patterns with real older adults in scenario‑based decisions?

\(\bullet\) Applying. How do the core dimensions of Putnam’s Social Capital Theory affect the adaptation difficulties of elderly groups in smart elderly care?

4.2 Agent and Simulation Configuration↩︎

We evaluate three LLMs to assess the generalizability of our framework. Qwen2.5-14B-Instruct [5] serves as the default agent engine, while additional results for GPT-4 [6] and GLM-4 [7] are provided in Appendix 11 (modeling task) & 13.3 (applying task). We set temperature = 0.7, top-p = 0.9. All key experiments are repeated five times; results are reported as means and standard errors. We generate 200 agents as detailed in Appendix 13.1 (data construction) & 13.2 (demographic profiles and behavioral analysis). For modeling task, we randomly sample 20 agents stratified by SES into low-, mid-, and high-SES groups for multi-round collective action experiments, each running for 25 rounds. For applying task, we use all 200 agents for smart elderly care simulations, also over 25 rounds.

4.3 Human Benchmarking↩︎

We conduct parallel experiments with 20 real older adults (\(age \geq 60\)) recruited online for the modeling task. Quota sampling matches the SES proportions of the 20 CGSS subsample. Each participant completed the same eight scenario-based decisions as the LLM agents (see Appendix 12.1 for details). Prior to task completion, participants provided informed consent and completed self‑report inventories on demographics, social trust, and norms, aligned with agent initialization. Results are presented in §5.3, further analyzed in Appendix 12.2.

5 Modeling Results: Theory Replication↩︎

In this section, we examine whether SocaSim can reproduce Putnam’s core mechanisms and align with human decisions. We present results from three complementary perspectives: macro-level pattern replication, ablation study, and human‑agent alignment validation.

5.1 Macro‑level Pattern Replication (Figure 3)↩︎

To examine whether the simulated macro-level patterns align with Putnam’s theoretical predictions, we analyze the core dynamics across four aspects, with results shown in Figure 3. Collective action outcomes are evaluated by the cooperation success rate, defined as the proportion of successful collective actions per round. In each round, agents interact within the social network and update their trust and norms based on historical experience.

\(\bullet\) Social Networks and Collective Action.

Figure 3 (a) demonstrates co‑evolution between network density, cooperation rate, and collective action success. Over 25 rounds, network density increased from 0.158 to 0.342, while cooperation rate and success rate rose to 0.795 and 0.762, respectively. Correlation analysis shows strong positive associations between network density and both collective action success (\(r \approx 0.976\), \(p < 0.001\)) and cooperation rate (\(r \approx 0.993\), \(p < 0.001\)). Results are stable across five runs with low variance.

\(\bullet\) Trust Evolution across SES Groups.

Figure 3 (b) shows the evolution of trust across different SES groups and trust types over 25 rounds. Trust within the same SES group consistently remains the highest, rising from 3.98 to 4.48, while trust toward distant SES groups steadily increases from 2.02 to 2.78. Both specific and general trust grow in parallel, from 2.82 to 3.68 and from 2.46 to 3.47, respectively. These upward trends are consistent across runs with narrow variance.

\(\bullet\) Norms and Cooperation Durability.

Figure 3 (c) illustrates the evolution of specific and general reciprocity norms over 25 rounds. Specific reciprocity norms rise from 2.92 to 3.78, and general reciprocity from 2.61 to 3.58, maintaining a parallel growth trajectory. The simulation produces 15 sustained cooperation pairs.

\(\bullet\) Bonding and Bridging Social Capital.

Figure 3 (d) compares the growth of bonding and bridging social capital over 25 rounds. Bonding capital increases from 2.78 to 6.95, whereas bridging capital grows from 1.91 to 4.92, resulting in a final gap of 2.03. We further quantify cross‑SES connections through the economic connectivity (EC) index, defined as the proportion of high‑SES friends in a low‑SES agent’s network, which rises steadily across rounds.

5.2 Ablation Study (Figure 4)↩︎

To investigate how each component contributes to the simulation performance of SocaSim, we conduct an ablation study to validate the necessity of each module, with results shown in Figure 4.

\(\bullet\) w/o SST. Overall averages change only slightly, but the structural patterns are largely removed. In particular, the distinction between same‑SES and distant groups collapses, with bonding/bridging capital shifting from 6.95/4.92 to 5.95/5.85.

\(\bullet\) w/o BDI. Cooperation rate drops by 71.6%, and general trust falls from 3.47 to 1.39.

\(\bullet\) w/o SCM. Sustained cooperative pairs decrease by 73.3%, and specific reciprocity drops by 17.5%.

5.3 Human Validation (Figure 5)↩︎

To validate group-level consistency between LLM agents and real older adults, we compare their adoption willingness across eight scenarios that vary social network density (high/low), trust level (high/low), and reciprocity norm (present/absent). The 20 real older adults were recruited as described in § 4.3. The results are summarized in Figure 5.

\(\bullet\) Overall alignment (Figure 5 (a)). Human and agent willingness rates are highly correlated (Pearson r = 0.974), with points tightly clustered around the identity line. Agents show slightly more extreme responses, higher under favorable conditions and lower under unfavorable ones.

\(\bullet\) Effect comparison (Figure 5 (b)). All coefficients are positive and significant (\(p < 0.01\)), and the ranking is identical for both groups, trust has the strongest effect, followed by reciprocity norm, and then social network density. Agents consistently produce larger coefficients, reflecting their cleaner causal reasoning that is less affected by noise such as emotion, fatigue, or social desirability, and thus showing more extreme responses.

\(\bullet\) Turnaround by SES (Figure 5 (c)). For each dimension, we measure the switch from non-adoption to adoption when it improves, while keeping the other two dimensions fixed, and stratify by SES. In both groups, switching rates rise with SES and trust dominates. For low-SES agents, the trust switching rate is 16.7% compared to 33.3% for humans, while the rates for social network and norms are both 0% versus 16.7% for humans. For high-SES agents, the trust switching rate reaches 83.3% compared to 66.7% for humans. These results indicate that LLM agents may respond more strongly than humans to SES-based structural inequality.

6 Applying Results: Elderly Care Study↩︎

In this section, we investigate how the three core dimensions of Putnam’s Social Capital Theory affect the adaptation challenges of elderly groups in smart elderly care. We present results from two complementary perspectives: scenario simulation and counterfactual intervention.

6.1 Scenario Simulation (Figure 6)↩︎

We model the adoption of a smart elderly care platform as a binary choice using all 200 agents. The simulation tracks three multidimensional metrics: technology adoption rate, defined as the proportion of agents that adopt the platform; psychological distress, measured via psychological pressure value and anxiety level; and decision contradiction, reflected by the frequency of adoption status switches. The results are shown in Figure 6.

\(\bullet\) Technology Adoption and SES.

Figure 6 (a) shows adoption rates across SES groups over 25 rounds. The high‑SES group rises steadily from 12.2% to 84.5% with minimal fluctuation. The low‑SES group increases slowly from 2.3% to 56.8%, with greater variability and wider confidence bands. The mid‑SES group stabilizes in the low‑70% range. All groups grow rapidly in the first 10 rounds, then slow with diminishing returns.

\(\bullet\) Psychological Pressure and Anxiety.

Figure 6 (b) illustrates the evolution of psychological pressure and anxiety. Both metrics rise steeply in early rounds (pressure from 0.10 to 0.27, anxiety from 0.12 to 0.26) and then stabilize. Pressure remains slightly higher than anxiety throughout, and variance widens modestly over time.

\(\bullet\) Decision Contradiction and Dynamics.

Figure 6 (c) presents the evolution of cumulative decision contradictions. The metric rises steadily from near 0 to about 50 by round 25. Its growth rate is consistent without clear inflection points. Variance across five runs diverges after round 10.

6.2 Counterfactual Intervention (Table 1)↩︎

To enhance causal interpretability and simulate a policy intervention, we design a counterfactual experiment. Since trust is the strongest dimension (Figure 5 (b)), we increase the initial trust of low‑SES agents by 1.0 (capped at 5.0) while keeping all other parameters unchanged, mimicking a targeted policy for disadvantaged groups.

We illustrate the effect using a case study (Figure 7). Before intervention, low trust causes hesitation, repeated decision contradictions, and eventual rejection. After trust is raised, the agent adopts stably, with anxiety and contradictions dropping sharply. This demonstrates trust’s causal role in stability. Table 1 compares the aggregate outcomes at round 25. After the intervention, technology adoption rises by 15.4%, pressure and anxiety fall by 19.8% and 22.4%, and decision contradictions decrease by 25.5%. These results indicate that boosting initial trust of low‑SES groups reduces the digital divide and improves decision stability.

7 Discussion↩︎

Our findings carry several important implications for policy, highlight the unique strengths and limitations of using LLM agents for social simulation, and open up new avenues for interdisciplinary research. For a more detailed discussion of these points, we refer the readers to Appendix 14.

8 Related Work↩︎

Putnam’s Social Capital Theory. Putnam’s Social Capital Theory defines social capital as rooted in social network, trust, and norms putnam1993?, putnam2000?. It serves as a core paradigm for explaining civic engagement, government performance, and democratic functioning tzanakis2013?, dodd2015?, and is widely applied to examine regional development annamalah2023?, public health carpiano2006?, and education attainment goddard2003?, careemdeen2021? as outcomes of resource accumulation within community structures. Recent studies further explore its role in crises chawa2023? and sustainable development sumi2025?. However, conventional empirical methods face practical constraints on experimental control and replication.

LLM-based Multi-agent Social Simulation. The rise of LLMs has enabled human-like agents capable of reasoning and social interaction wang2026?, sudhakar2025?, wang20251?. These agents perform complex social tasks zhang20251?, xu2025?, leading to growing use of LLM-based social simulation across scenarios piao2025?, yang2025?, zhang2025?. For example, generative agents piao2025? simulate intricate social behaviors; ProSIM [8] models prosocial behavior; and SOTOPIA-\(\Omega\) zhang2025? integrates human negotiation strategies. However, existing frameworks have not yet provided a reproducible and controllable environment for systematically instantiating Putnam’s Social Capital Theory, nor have they offered process-level interpretability of how trust accumulates or norms are internalized. Our work aims to address this gap.

9 Conclusion↩︎

We presented SocaSim, a novel LLM‑based multi‑agent simulation framework to model and apply Putnam’s Social Capital Theory. By integrating social network evolution, trust dynamics, and norm propagation into a unified environment, we analyzed collective‑action behaviors and extended the framework to smart elderly care. Our results demonstrate a consistent alignment between agent behaviors and human social dynamics, revealing how social capital shapes group coordination and technology adoption. Moreover, SocaSim offers a scalable, reproducible approach to investigate social mechanisms across social science domains.

Limitations↩︎

In this section, we discuss the limitations of our work as follow:

• Our current framework models the dynamic evolution of social capital primarily through text-based interactions, whereas in reality, trust building and norm propagation often rely on multimodal signals such as speech and images. Incorporating multimodal information is an important direction to enhance simulation realism, but it also brings challenges in data acquisition, cross-modal alignment, and ethics. Therefore, exploring multimodal social interaction mechanisms is a highly valuable research direction for future work.

• Due to recruitment costs and resource constraints, the number of participants in our human alignment experiment is limited, which may affect the statistical generalizability of the findings. Future work should increase the sample size and include more diverse populations to more robustly analyze alignment between simulated and real behaviors across different demographic dimensions.

• Our study is based on CGSS data and has not incorporated comparative analyses across diverse cultural contexts. The accumulation and dynamics of social capital may be shaped by cultural values, institutional settings, and related factors. Cross-cultural comparisons are thus a promising direction to evaluate both generalizability and context-specificity.

Ethics Statement↩︎

As the use of LLMs for social simulation grows, it is crucial to consider the ethical implications of deploying such systems to model social capital dynamics and elderly care behaviors. While this work explores the potential of LLM‑based multi-agent simulations to study Putnam’s Social Capital Theory, all experiments are conducted solely within controlled research settings and remain an early-stage theoretical exploration.

The SocaSim framework and all simulations are designed for scientific theory validation and methodological development rather than for direct real‑world decision making, policy formulation, or replacing the complexity of human social systems. We strongly emphasize that any potential application of our findings to smart elderly care policy, community governance, or service design must involve human expert oversight, ethical review, and careful consideration of fairness to mitigate risks arising from model simplifications or data biases.

All demographic data used for agent initialization are drawn from the 2023 CGSS elderly subsample, which is released as publicly available aggregate statistics and contains no personally identifiable information. The construction of synthetic agents does not involve direct simulation or identity mapping of real individuals. For our human‑agent alignment validation, 20 real older volunteers participated with informed consent. The study consisted solely of anonymous questionnaires and scenario-based decision tasks, with no interventional or deceptive procedures. Participants’ personal information was used only for SES stratification and anonymized analysis, and they received reasonable compensation in accordance with local ethical guidelines. The research protocol was approved by the institutional ethics board.

Looking forward, we will continue to focus on fairness, transparency, and accountability in agent-based social simulation. We will actively expand cross-cultural and multimodal simulation capabilities, and develop human-in-the-loop experimental paradigms. We remain committed to advancing this field toward responsible, inclusive, and socially beneficial directions.

10 Supplementary Preliminary↩︎

Putnam’s Social Capital Theory putnam1993?, putnam2000? offers a theoretical blueprint of how social network, trust, and norms interact in self‑reinforcing ways [9], [10], and it has influenced work in fields ranging from political science and public administration to economic development mikiewicz2021?, zhou2023?, thang2025?. This section further explains these three dimensions, as illustrated in Figure 8, by showing how they jointly shape a group’s ability to overcome collective‑action dilemmas. We also provide concrete positive and negative examples for each dimension, and highlight how the interplay among these dimensions affects coordination, responses to shared problems, and sustained participation.

1. Social Network Dimension (Figure 8 (a))↩︎

Putnam highlights the importance of civic and associational networks in enabling coordination and mutual support putnam1993?. In this sense, the social network dimension concerns the existence, density, and stability of social ties, which determine how easily information, resources, and assistance can circulate within a group.

10.0.0.1 Positive example (strong connectivity network).

In a community where people know each other well and interact frequently, a proposal for a joint project spreads quickly through trusted channels. Members readily accept, offer help, and recruit collaborators. As a result, coordination is smooth, and the project succeeds with minimal delays.

10.0.0.2 Negative example (weak connectivity network).

In a fragmented community where residents rarely interact, a proposal for a joint project is met with suspicion. Some members prefer to act alone, believing others are unreliable, while others assume their neighbors are plotting behind their backs. Consequently, cooperation fails, the project stalls, and potential benefits are lost.

2. Trust Dimension (Figure 8 (b))↩︎

Trust is another central component of social capital. Putnam argues that trust facilitates coordination by reducing uncertainty about others’ behavior and lowering the costs of collective action putnam1993?. When trust is present, individuals are more willing to rely on one another and to participate without monitoring or verification.

10.0.0.3 Positive example (high-trust scenario).

In a group where members have consistently kept their promises, a new collaborative task is met with immediate and full commitment. People assume others will contribute fairly, so they freely invest effort without hesitation, openly share information, and complete the task efficiently.

10.0.0.4 Negative example (low-trust scenario).

In a group where past experiences have been marked by broken commitments, even a well‑designed proposal is met with suspicion. Members insist on checking each other’s reliability, avoid committing, and prefer to watch how things unfold. They spend excessive time verifying intentions and drafting detailed contracts. As a result, cooperation is delayed, costs rise, and the goals are often not reached.

3. Norms Dimension (Figure 8 (c))↩︎

Putnam also emphasizes the importance of shared norms, especially norms of reciprocity putnam2000?. This dimension concerns widely accepted expectations about appropriate behavior, such as mutual support, reciprocal help, and continued participation. Strong norms make sustained collective action more likely in situations where immediate returns are uncertain or delayed.

10.0.0.5 Positive example (strong norm context).

In a community where mutual aid is taken for granted, a long‑term infrastructure project that will only pay off after many months receives broad support from the start. Members openly affirm that helping each other is how things are done, and they trust that sticking to the mutual‑help rule will eventually benefit everyone. They contribute without expecting an immediate reward, and the project proceeds steadily to successful completion.

10.0.0.6 Negative example (weak norm context).

In a community where people primarily look after their own short‑term interests, the same long‑term project meets immediate resistance. Each potential participant focuses on immediate personal gain and refuses to join without an instant payoff. Some even suggest abandoning the project altogether, while others prefer to wait until a clearer reward appears. As a result, the project never gets off the ground, and the prevailing attitude is that helping each other is not a shared expectation.

This theoretical logic directly supports our study in two respects. First, it provides a framework for formalizing how social capital shapes a group’s capacity to overcome collective-action dilemmas, which forms the basis of our theoretical modeling task. Second, it offers an analytical lens for smart elderly care, where adaptation depends not only on individual capability, but also on the surrounding structure of social relationships, trust, and supportive norms. This motivates our applying task, in which we examine how the three dimensions of social capital are related to the adaptation dilemmas faced by older adults in smart elderly care.

11 Cross‑Model Consistency in Macro‑level Pattern Replication↩︎

To address reproducibility concerns and validate the generalizability of our findings, we conducted additional experiments using three LLMs with distinct architectures: Qwen‑2.5‑14B[5], GPT‑4[6], and GLM‑4[7]. The results demonstrate consistent replication of the macro‑level patterns of Putnam’s Social Capital Theory across its three core dimensions (social network, trust, and norms) and two forms of social capital (bonding and bridging), based on data from the final round of simulation. All models reproduce core patterns aligned with Putnam’s Social Capital Theory.

11.0.0.1 \(\bullet\) Social Network Dimension (Table ¿tbl:table3?):

Every model exhibits high network density, cooperation rate, collective action rate, and help rate, consistent with the macro‑level pattern that dense networks facilitate cooperation and mutual aid.

11.0.0.2 \(\bullet\) Trust Dimension (Table [tbl:table4]):

The results show that specific trust, general trust, and group trust (same SES) are higher than group trust (distant SES), replicating the structural nature of trust.

11.0.0.3 \(\bullet\) Norms Dimension (Table [tbl:table5]):

High rates of specific and general reciprocity, along with sustained cooperation pairs, indicate that norms are maintained and reinforced through long‑term interaction, reproducing the macro‑level patterns.

11.0.0.4 \(\bullet\) Social Capital Accumulation (Table [tbl:table6]):

All models exhibit growth in both bonding and bridging capital by the end of the simulation, confirming the dynamic accumulation of social capital as a macro‑level phenomenon.

All reported metrics are averaged over five independent runs with standard deviations shown. The consistent trends across models underscore the robustness and generalizability of our macro‑level pattern replication.

12 Human Experiment Details↩︎

12.1 Experimental Procedure↩︎

We recruited 20 real older adults online to participate in the modeling task. Given the digital literacy barriers among some older adults, the experiment was designed as a simple questionnaire rather than a complex online platform. Each participant was presented with the same fixed scenario background followed by eight scenario‑based decisions in random order. For each scenario, participants were asked to choose between two options: willing or unwilling. The questionnaire also collected demographic information, self‑reported social trust, and reciprocity norms. The fixed scenario background and the eight scenarios are illustrated in Figure 9.

Prior to the experiment, all participants provided written informed consent and were clearly informed about the purpose of the study, the voluntary nature of their participation, and the measures taken to protect their privacy and data. The entire session took approximately 15 minutes to complete, and each participant received 20 CNY, equivalent to 80 CNY per hour, which is comparable to the typical hourly rate for simple online tasks in China.

Each LLM agent received a prompt whose scenario, question, and response format were designed to be completely identical to those used in the human survey. The community‑level social variables, namely social network density (high/low), trust level (high/low), and reciprocity norm (present/absent), were fully crossed in a factorial design, resulting in eight distinct scenarios. The presentation order of the eight scenarios was randomized for each agent. The prompt template and the eight community scenarios are illustrated in Figure 10. In addition, Table [tab:agent-profiles] presents the 20 agents that were randomly drawn from the 200 elderly subsample for the modeling task, covering low-SES, mid-SES, and high-SES strata.

12.2 Analysis: Divergence Between Simulated and Real older Adults↩︎

Based on the human alignment experiments in the modeling task, we further analyze the similarities and differences between real older adults and LLM agents, as summarized in Figure 11.

13 Smart Elderly Care↩︎

13.1 Dataset Generation↩︎

As shown in Figure 12, the CGSS 2023 dataset is processed through a structured pipeline to construct agent profiles for social simulation. Key variables are transformed sequentially: age is derived from birth year and filtered to elderly adults (60–100 years); gender is mapped to categorical labels; education is converted into both category labels and standardized years of schooling; family income is cleaned and percentile-ranked; self‑rated health status is translated into a label and a normalized score; occupation is probabilistically assigned based on age and education; and general trust is reverse‑coded. Finally, we compute a SES index as a weighted composite of education years, normalized income, and occupation following standard sociological practice, and classify agents into low-SES, mid-SES, and high-SES groups. This end‑to‑end pipeline produces demographically grounded and methodologically consistent profiles suitable for agent‑based modeling.

13.2 Agent Profiles and Behavioral Analysis↩︎

Following the data processing pipeline, we generated 200 agent profiles tailored for the smart elderly care scenario, which are subsequently instantiated as interactive agents in the simulation. Figure 13 presents the demographic distributions of this elderly agent population across age, gender, education, income, health, and SES dimensions.

Table ¿tbl:table7? reports the significance tests of technology adoption rates across different demographic attributes. The analysis shows that age, gender, education, income, health, and SES all exert significant influences on adoption behavior (\(p < 0.05\)). Specifically, younger (60–64 years), male, highly educated (college or above), higher‑income, healthier, and high‑SES elderly agents show significantly higher adoption rates, which confirms that the digital divide persists within the elderly population and is shaped by multiple socioeconomic factors.

To further uncover behavioral heterogeneity, Table [tbl:table8] and Table 2 detail cooperative helping behaviors and technology adoption patterns across demographic subgroups:

\(\bullet\) Cooperation and mutual help (Table [tbl:table8]): Younger, better educated, higher‑income, and higher‑SES agents consistently exhibit higher cooperation rates, help rates, and greater trust growth. This shows that the accumulation of social capital, especially reciprocity and trust, is closely tied to demographic and socioeconomic structure.

\(\bullet\) Technology‑adoption behaviors (Table 2): Advancing age is associated with lower adoption rates, elevated anxiety levels, and more frequent decision reversals. In contrast, higher education and higher SES correspond to higher adoption, lower anxiety, and more stable decisions. These fine‑grained patterns highlight both the psychological and behavioral barriers faced by older adults in technology adoption, offering an empirical foundation for designing stratified, context‑sensitive support strategies in smart elderly care.

13.3 Analysis of Technology Adoption and Psychological Impact↩︎

To verify the robustness and generality of the simulation results, we conducted parallel experiments using three LLM architectures (Qwen‑2.5‑14B, GPT‑4, and GLM‑4), analyzing data from the final simulation round. The results show that the technology‑adoption behavior and psychological indicators of the agents exhibit a consistent socioeconomic stratification pattern across models, with overall trends aligning with the single‑model analysis presented earlier.

Regarding technology adoption rates (Table ¿tbl:table10?), all three models reveal a clear SES gap (\(SES\) \(Gap \approx 0.27\)), with the high‑SES group consistently showing significantly higher adoption rates than the low‑SES group. Although overall adoption rates vary by model (highest for Qwen2.5‑14B, relatively lower for GLM‑4), the stable influence of SES on technology uptake remains consistent across models, further supporting Putnam’s thesis that social capital structurally facilitates adoption among high‑SES groups while increasing uncertainty among low‑SES groups.

For psychological metrics (Table [tbl:table11]), as the model changes from Qwen2.5‑14B to GPT‑4 to GLM‑4, the agents show increasing levels of mental pressure, anxiety, and decision contradictions. Agents simulated with GLM‑4 show the highest psychological pressure (\(0.295\pm0.042\)) and the most frequent decision reversals (\(60\pm12\)), indicating that the response traits of the model itself may indirectly shape the agents’ mental burden and decision stability. This result is consistent with the evolving trends of pressure and anxiety shown in Figure 6 (b) and further reveals that during technology adoption, low‑SES groups with weaker social‑support networks are more prone to higher decision volatility due to adaptation stress.

14 Supplementary Discussion↩︎

14.1 Policy Implications↩︎

Our simulations reveal a clear socioeconomic gradient in technology adoption among older adults. Low-SES groups not only adopt smart elderly care platforms at lower rates but also experience higher psychological pressure and more frequent decision contradictions. This pattern indicates that access to digital services alone may not be sufficient to narrow the digital divide. In our setting, trust plays a central role in shaping adoption outcomes, suggesting that policies which strengthen trust-related resources could help mitigate structural disparities.

These observations point to several policy directions worth consideration. First, community trust can be fostered by involving trusted local institutions, such as community health centers or residents’ committees, in recommending platforms. This may lower the initial barrier for low-trust groups by providing credible endorsements and reducing perceived uncertainty [11]. Second, peer effects can be leveraged by highlighting successful adoption among high-SES users, which can spread norms through a peer-influence ripple effect (e.g., "others are using it, so I will try too") [12]. Third, more inclusive services that embed psychological support into the adoption process, for instance through step-by-step guidance and reassurance after failures, can alleviate adaptation stress [13].

Taken together, these complementary measures point toward shifting technology adoption from an isolated individual decision to a more collective, trust-supported learning process.

14.2 Boundaries↩︎

Our experiments show that LLM agents have both unique strengths and clear limitations when reproducing social capital dynamics.

The main advantage lies in process‑level interpretability. The round‑by‑round simulation tracks how interactions update social network, trust, and norms over time, making it possible to see dynamics such as the gradual build‑up of trust and its rapid collapse after a single failure. These are patterns that standard statistical analyses often compress into aggregate summaries instead of revealing step by step. A second benefit is a clearer isolation of mechanisms. By keeping transient human factors such as short‑term emotion, fatigue, and social‑desirability pressures roughly constant, agents display more pronounced causal responses, becoming more proactive under favorable conditions and more passive under unfavorable ones. This makes the differential effects of trust, norms, and social networks easier to detect and the causal chain easier to follow, although this added clarity comes at the cost of reduced human realism, a trade‑off we examine in the limitations below.

The limitations are equally important. Agents have lower behavioral heterogeneity than humans, with response variability only one‑twentieth of that observed in real populations. They are also less sensitive to abstract social norms such as a general tradition of mutual help, compared to explicit trust signals like a recommendation from a known authority. Moreover, agents find it difficult to simulate the emotional fluctuations and intent attribution that occur in long‑term human interactions. Therefore, while LLM agents are effective tools for studying central tendencies, main effects, and theoretical mechanisms, research that involves behavioral heterogeneity, emotion‑driven processes, or culturally embedded social behaviors must still rely on real human subjects.

14.3 Interdisciplinary Research↩︎

On the surface, social scientists and computer scientists both care about the relationship between language and behavior, but their motivations are fundamentally different. Social scientists use language to understand human thoughts and feelings, whereas computer scientists use language to predict behavior [14]. This study demonstrates a hybrid paradigm that bridges the two. We use the computability of LLM agents to simulate the dynamic coupling of social capital, and at the same time we validate the model’s psychological and sociological credibility through alignment with real human behavior. This bidirectional interplay of understanding and prediction creates a shared platform for both disciplines.

15 Prompts↩︎

15.1 Prompt for LLM-based Agents↩︎

You are playing the role of an elderly person in a community that is gradually adopting smart elderly care technologies. Fully immerse yourself in the following identity.

Demographics:
- Age: {age}
- Gender: {gender}
- Health Status: {health_status}
- Education Level: {education}
- Monthly Income: {income} CNY
- Occupation: {occupation}
- Socioeconomic Status: {ses_category}

Social Capital Attributes:
- Trust Level toward Specific Individuals: {trust_specific}/5.0
- Trust Level toward the Broader Community (General Trust): {trust_general}/5.0
- Reciprocity Norm Strength: {norm_general}/5.0
- Social Capital Type Preference: {capital_preference} (mainly Bonding, mainly Bridging, or Balanced)
- Technology Experience: {tech_experience}

Current Situation:
{situation_description}

Based on your demographic background, social capital attributes, and the current situation, what action would you take?
Before deciding, briefly think through:
- How do your age, health, SES, and tech experience shape your view of this situation?
- How does your trust tendency (both specific and general) influence your willingness to engage?
- What do you most want right now: concrete help, social connection, risk avoidance, or maintaining relationships?

Please respond with your decision and a brief first-person reasoning.

You have just completed a round of interaction with another community member. Reflect on the outcome and describe how your social state has changed.

Interaction Record:
- Partner: {partner_name} (Relationship: {relationship_type}, SES: {partner_ses})
- What was promised: {promised_action}
- Actual outcome: {actual_outcome} (choose from: "Fully fulfilled", "Partially fulfilled", or "Not fulfilled / Betrayal")

Your Social State Before This Interaction:
- Specific Trust toward {partner_name}: {old_specific_trust}/5.0
- General Trust in the community: {old_general_trust}/5.0
- Reciprocity Norm Strength: {old_norm}/5.0

[Interaction History Context]
- Consecutive rounds without any reciprocity experience: {no_reciprocity_count}
- Your preferred social capital type: {capital_preference} (Bonding-oriented, Bridging-oriented, or Balanced)
- Interaction type of this round: {interaction_type} (strong-tie / cross-group weak-tie)

Please provide a first-person narrative (about 60 words) that describes:
- How did this interaction make you feel compared to before?
- Has your trust in this partner increased, decreased, or stayed the same? Why?
- Has your faith in community reciprocity strengthened or weakened? Why?
- Did this round strengthen your emotional bond with your in-group, or help you build a bridge to a different group?
- Briefly indicate what this means for your willingness to cooperate in the next round.

Your description will be used to update your internal state. Be honest and consistent with your profile and past experiences.
Example Output: "Mr. Li did exactly what he said he would, which made me feel relieved. My trust in him has grown a little because he has been reliable for several rounds now. I still feel a bit distant from the broader community, but among my close neighbors, I feel safe and willing to help again next time."

At the end of this round, reflect on your recent experiences and summarize your current state of mind.

Summary of Recent Interactions:
{interaction_history_summary}

Your Current Social State:
- General Trust in the community: {general_trust}/5.0
- Reciprocity Norm Strength: {norm_general}/5.0

Please write a first-person summary (about 50 words) that covers:
- The most important thing you learned or felt during this round.
- How your overall trust in people and your belief in mutual help have shifted, if at all.
- What kind of social strategy you intend to follow in the next round (e.g., stick to close friends, cautiously try new connections, keep distance from unreliable people).

Example Output: "This round reminded me that not everyone in the community keeps their word. I was a bit disappointed by one person's empty promise, but my close neighbors are still dependable. Next time I will be more selective about whom I cooperate with, and I will rely more on people I already know well."

15.2 Prompt for Cooperation Decision↩︎

You are an agent with the following profile:
- SES: {ses_category}
- Specific Trust toward this Partner: {trust_level}/5.0
- General Trust in Community: {general_trust}/5.0
- Reciprocity Norm Strength: {norm_strength}/5.0

Another agent (relationship: {relationship_type}, SES: {partner_ses}) proposes a cooperation:
- Proposal: {proposal_description}
- Your required effort: {effort_level}
- Expected benefit if cooperation succeeds: {benefit}

[Phase 1: Proposal Response]
Based on your characteristics, would you cooperate?
Before deciding, consider:
- Belief: How reliable is this partner given your past interactions and their SES group?
- Desire: Do you prioritize maintaining trust, gaining benefit, or avoiding risk?
- Intention: Make your decision.

Answer with: ACCEPT or REJECT, followed by a brief reason.

[Phase 2: Execution Fulfillment -- only if you accepted above]
You previously accepted this cooperation. Now decide privately whether to honor your commitment.
- Your promised action: {promised_action}
- Immediate gain if you defect: {defect_gain}
- Other participants likely to honor: {expected_cooperation_rate}
- Rounds remaining: {remaining_rounds}

Consider:
- Belief: Will your partner also honor? Does your reputation matter at this stage?
- Desire: Immediate gain vs. long-term trust vs. guilt avoidance?
- Intention: Decide to FULLY HONOR, PARTIALLY FULFILL, or NOT HONOR.

Answer with: COOPERATE or DEFECT, followed by a brief reason.

15.3 Prompt for Technology Adoption Decision↩︎

You are an elderly person considering whether to start using a new smart elderly care platform.

Your Profile:
- Age: {age}
- Health Status: {health_status}
- Socioeconomic Status: {ses_category}
- Technology Experience: {tech_experience}
- Platform Trust (how much you trust this technology provider): {platform_trust}/5.0
- General Trust in Community: {general_trust}/5.0
- Reciprocity Norm Strength: {norm_general}/5.0

Technology Information:
- Type: Smart elderly care platform
- Benefits: Health monitoring, emergency assistance, social connection
- Perceived Risks: Privacy concerns, technical complexity, monetary cost

Social Influence:
- Close friends/neighbors already using it: {friends_using}/{total_friends}
- Family members recommend it: {family_recommends} (Yes/No)

Would you adopt this technology?
Before deciding, reason step-by-step:
- Belief: How trustworthy is the platform and the people recommending it? What do you make of the risks?
- Desire: Do you want better health security, connection with others, or to avoid the burden of learning new tech?
- Intention: Weigh your SES, health needs, social circle, and family advice to make a final choice.

Answer with: ADOPT or REJECT, followed by a brief first-person reason.

References↩︎