Pith. sign in

REVIEW 3 major objections 7 minor 42 references

LLM agents reproduce social capital theory and trace trust's causal role

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · glm-5.2

2026-07-08 17:07 UTC pith:PL74S4BR

load-bearing objection Real application of LLM agents to Putnam's theory, but the central causal claim is near-tautological due to prompt architecture the 3 major comments →

arxiv 2607.06080 v1 pith:PL74S4BR submitted 2026-07-07 cs.CL cs.AIcs.SI

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

Shiyi Ling , Zhi Zheng , Hui Zheng , Wenjun Xue , Feng Ye , Tong Xu This is my paper
classification cs.CL cs.AIcs.SI
keywords social capital theoryLLM agentsmulti-agent simulationtrust dynamicsnorm propagationsocial network evolutioncollective actionsmart elderly care
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper introduces SocaSim, a multi-agent simulation framework where LLM-driven agents endowed with demographic and social-capital attributes interact over repeated rounds to model Putnam's Social Capital Theory. The framework integrates three dynamically coupled dimensions: social network evolution, trust adaptation, and norm propagation. Each agent carries a socioeconomic profile (drawn from the Chinese General Social Survey), makes decisions through a Belief-Desire-Intention reasoning module, and updates its cognitive state via a Social Cognitive Memory that tracks interaction history and adjusts trust and norm levels. The simulation runs in two-phase rounds: a Proposal phase where agents initiate or respond to collective actions, and an Execution phase where they privately decide whether to honor commitments. The authors claim that this setup reproduces the macro-level patterns Putnam predicted, including the self-reinforcing feedback between network density and cooperation, the SES-stratified accumulation of trust, and the durability of cooperation through internalized reciprocity norms. They further claim strong group-level alignment with human data, reporting a Pearson correlation of 0.974 between agent and older-adult decisions across eight scenario conditions. In a counterfactual intervention, boosting low-SES agents' initial trust by one point raised technology adoption by 15.4 percent and reduced decision contradictions by 25.5 percent, which the authors interpret as evidence that trust functions as a causal lever for reducing the digital divide.

Core claim

The central object is SocaSim itself, an LLM-based multi-agent framework whose architecture rests on three modules: a Social Structure Trait (SST) module that encodes demographic and socioeconomic attributes into each agent's prompt, a Belief-Desire-Intention (BDI) module that drives context-sensitive decision-making, and a Social Cognitive Memory (SCM) module that handles adaptive learning through multi-layer trust updates, norm formation and decay, and reflective memory. The paper's core claim is that these three modules, running in a two-phase Proposal-Execution round structure, are sufficient to reproduce the dynamic coupling of social networks, trust, and norms that Putnam's theory pos.

What carries the argument

SocaSim framework: SST module (demographic/socioeconomic encoding), BDI module (LLM-based reasoning), SCM module (trust/norm/memory updates), two-phase round structure (Proposal then Execution), 200 agents initialized from CGSS 2023 elderly subsample, 25-round simulations, counterfactual trust intervention

Load-bearing premise

The load-bearing assumption is that LLM agents' decisions, driven by natural-language prompts encoding demographic and social-capital attributes, faithfully approximate human social decision-making processes. If LLM outputs are prompt-elicited pattern completions shaped by the model's training exposure to sociological concepts rather than genuine social reasoning, the simulation's causal claims about trust, norms, and adoption are artifacts of prompt design rather than证据about

What would settle it

Run the same eight-scenario human-alignment test with a substantially larger and more diverse participant pool (e.g., 200+ participants across multiple regions and cultural contexts). If the Pearson correlation drops substantially below 0.974, or if the coefficient ranking (trust > norms > network) no longer matches, the group-level alignment claim weakens. Additionally, vary the prompt formulations for social-capital attributes (e.g., rephrase 'low trust' using different language) and test whether the simulation's macro-level patterns remain stable; if they shift significantly with prompt ph

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If LLM agents can faithfully model social capital dynamics, researchers gain a low-cost, fully controllable testbed for theories that are otherwise difficult to study through surveys or field experiments, enabling repeated replication and systematic parameter variation.
  • The counterfactual intervention result, where boosting low-SES trust raised adoption by 15.4 percent, suggests a concrete policy hypothesis: community trust-building programs (e.g., endorsements from trusted local institutions) could be a more effective lever for technology adoption among disadvantaged elderly than direct subsidies or technical training.
  • The ablation study's finding that removing the BDI module collapses cooperation (71.6 percent drop) while removing the SST module flattens inter-group structural differences provides a modular decomposition of which social-capital mechanisms matter most for which outcomes.
  • The framework's round-by-round tracing of trust accumulation and norm internalization offers a process-level view that aggregate statistical methods cannot produce, potentially allowing researchers to identify tipping points or critical thresholds in social dynamics.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The r=0.974 human-agent alignment is computed over only 8 scenario-level data points from 20 participants, meaning the correlation reflects group-level averages across conditions rather than individual-level behavioral fidelity. A framework could align perfectly at this aggregate level while still misrepresenting individual decision processes.
  • If LLM agents produce 'more extreme responses' and 'lower behavioral heterogeneity' than humans (as the paper itself notes), the counterfactual intervention results may overstate the magnitude of trust's causal effect relative to what a real-world policy would achieve, since the agents' cleaner causal reasoning amplifies structural signals.
  • The framework's reliance on natural-language prompts to encode social-capital attributes means the simulation's behavior is shaped by how the LLM interprets sociological constructs in text, not by independent measurement of those constructs. Two different prompt formulations of 'low trust' could produce different dynamics.
  • Cross-model consistency across Qwen-2.5-14B, GPT-4, and GLM-4 is presented as evidence of robustness, but all three models share training-data overlap with sociological literature, so convergence may reflect shared exposure to Putnam's theory rather than independent validation of the simulation's fidelity.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. The manuscript introduces SocaSim, an LLM-based multi-agent simulation framework designed to model Putnam's Social Capital Theory by integrating social network evolution, trust dynamics, and norm propagation. The framework uses agents initialized with demographic profiles from the Chinese General Social Survey (CGSS) and employs a Belief-Desire-Intention (BDI) architecture with Social Cognitive Memory (SCM) to simulate round-by-round collective action and technology adoption decisions. The authors evaluate SocaSim through macro-level pattern replication, an ablation study, human-agent alignment validation, and a counterfactual intervention applied to smart elderly care technology adoption. The central claims are that the simulation reproduces Putnam's theoretical patterns, achieves strong group-level alignment with human data (Pearson r=0.974), and enables process-level causal tracing of how trust influences adoption decisions.

Significance. The paper tackles an ambitious interdisciplinary problem: bridging social science theory (Putnam's Social Capital Theory) with LLM-based multi-agent simulation. The integration of SST, BDI, and SCM modules into a unified framework is a reasonable architectural contribution. The human-agent alignment study with 20 real older adults and the cross-model consistency checks (Qwen2.5-14B, GPT-4, GLM-4) add empirical grounding. The application to smart elderly care provides a concrete, policy-relevant test case. However, the significance of the causal claims is tempered by the prompt architecture, which I discuss below.

major comments (3)
  1. §6.2, Table 1, and Appendix F.3: The counterfactual intervention raises low-SES agents' initial trust by 1.0 and observes a 15.4% adoption increase. The paper frames this as evidence that 'trust can serve as a key causal lever.' However, the technology adoption decision prompt (Appendix F.3) explicitly provides 'Platform Trust: {platform_trust}/5.0' and 'General Trust in Community: {general_trust}/5.0' as direct numeric inputs to the LLM, and instructs the agent to reason about 'How trustworthy is the platform and the people recommending it?' before outputting ADOPT or REJECT. This means the 15.4% increase may largely reflect the LLM's sensitivity to a directly manipulated prompt parameter rather than an emergent causal property of the simulated social system. The paper needs to either (a) demonstrate that trust also influences adoption through emergent, non-prompt-mediated pathways (e.g
  2. §5.3, Figure 5(a): The human-agent alignment is computed as Pearson r=0.974 over only 8 data points, each representing a scenario-level average from 20 participants. With n=8, this correlation has very limited statistical power and is sensitive to outliers. The paper should report confidence intervals for this correlation and acknowledge the statistical thinness more prominently. Additionally, the 8 scenarios explicitly vary trust level (high/low) as a factorial dimension (Appendix C.1, Table 6), so both humans and agents are responding to the same manipulated variable. High correlation under these conditions is expected by construction to some degree, and the paper should discuss this more carefully.
  3. §5.1, Figure 3(a): The reported correlation between network density and collective action success (r≈0.976) is likely inflated by shared time trends. Both network density and cooperation rate increase monotonically over the 25 rounds, so the high correlation may reflect co-evolution along a common trajectory rather than a meaningful structural relationship. The paper should consider differencing or detrending the time series before computing correlations, or at minimum acknowledge this concern.
minor comments (7)
  1. §3.2.3: The norm decay mechanism triggers when 'no reciprocity is observed over ten consecutive rounds,' but the simulation runs for only 25 rounds with 20 agents. This threshold seems high relative to the simulation length and population size. Please justify this parameter choice.
  2. Table 6 in Appendix C.1 and Figure 9 appear to contain the same scenario information. Consider consolidating to avoid redundancy.
  3. §4.2: The paper states temperature=0.7 and top-p=0.9, but does not discuss how sensitive the results are to these sampling parameters. A brief sensitivity analysis or justification would strengthen reproducibility.
  4. Figure 4 (ablation study): The y-axis labels and metric definitions are unclear. Please clarify what metrics are being compared across the ablation conditions.
  5. §5.3, Figure 5(b): The paper attributes agents' larger coefficients to 'cleaner causal reasoning that is less affected by noise.' This is a strong interpretive claim. Consider softening to a more neutral description of the observed difference.
  6. The Economic Connectivity (EC) index is introduced in §5.1 but its formula is not provided. Please include the precise definition.
  7. Several references appear to have future dates (e.g., Piao et al., 2026; Wang et al., 2026; Zhou et al., 2026). Please verify these are correct.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a careful and constructive review. The three major comments each identify legitimate methodological concerns regarding (1) the causal interpretation of the counterfactual intervention given the prompt architecture, (2) the statistical thinness of the human-agent alignment correlation, and (3) potential inflation of the network density–cooperation correlation by shared time trends. We address each point below and commit to concrete revisions in all three cases.

read point-by-point responses

  1. Referee: §6.2, Table 1, and Appendix F.3: The counterfactual intervention raises low-SES agents' initial trust by 1.0 and observes a 15.4% adoption increase. The paper frames this as evidence that 'trust can serve as a key causal lever.' However, the technology adoption decision prompt (Appendix F.3) explicitly provides 'Platform Trust: {platform_trust}/5.0' and 'General Trust in Community: {general_trust}/5.0' as direct numeric inputs to the LLM, and instructs the agent to reason about 'How trustworthy is the platform and the people recommending it?' before outputting ADOPT or REJECT. This means the 15.4% increase may largely reflect the LLM's sensitivity to a directly manipulated prompt parameter rather than an emergent causal property of the simulated social system. The paper needs to either (a) demonstrate that trust also influences adoption through emergent, non-prompt-mediated pathways, or,

    Authors: The referee raises a valid and important concern. We agree that the current prompt architecture (Appendix F.3) provides trust as a direct numeric input to the LLM, which means the counterfactual intervention conflates two distinct channels: (1) the LLM's direct sensitivity to a manipulated prompt parameter, and (2) emergent, system-mediated pathways through which trust—accumulated via the SCM module through interaction history—affects downstream adoption decisions (e.g., via social influence from trusted neighbors who have already adopted, or via reduced decision contradictions that stabilize behavior over rounds). In the current manuscript, we do not adequately distinguish these two channels, and the causal claim 'trust can serve as a key causal lever' is therefore stronger than what the evidence supports. We will revise the manuscript in the following ways: First, we will soften the causal framing in §6.2 and the abstract to explicitly acknowledge that the observed 15.4% increase reflects both prompt-mediated and potentially emergent pathways, and that disentangling them requires further experimentation. Second, we will add a discussion of this limitation to the Limitations section. Third, we will outline as future work a controlled experiment in which trust is manipulated only through the interaction history (SCM) pathway—by, for example, injecting positive or negative interaction events—without directly modifying the numeric trust parameter in the adoption prompt, which would allow isolation of the emergent causal pathway. We note that the case study in Figure 7 does provide some qualitative evidence of emergent pathways: the agent's reasoning references her neighbor's adoption and family encouragement as factors that 'now carry more weight' after trust is raised,暗示 revision: partial

  2. Referee: §5.3, Figure 5(a): The human-agent alignment is computed as Pearson r=0.974 over only 8 data points, each representing a scenario-level average from 20 participants. With n=8, this correlation has very limited statistical power and is sensitive to outliers. The paper should report confidence intervals for this correlation and acknowledge the statistical thinness more prominently. Additionally, the 8 scenarios explicitly vary trust level (high/low) as a factorial dimension (Appendix C.1, Table 6), so both humans and agents are responding to the same manipulated variable. High correlation under these conditions is expected by construction to some degree, and the paper should discuss this more carefully.

    Authors: The referee is correct on both points. With n=8, the Pearson r=0.974 has a 95% confidence interval of approximately [0.85, 0.99], which is wide despite the high point estimate. We will report this confidence interval in the revised manuscript. We also agree that because the 8 scenarios are constructed by fully crossing three binary factors (network density, trust, norms), high correlation between human and agent responses is partly expected by construction: both groups are responding to the same systematic variation in social capital dimensions. The correlation demonstrates that agents and humans respond similarly to the same factorial manipulations, but it does not by itself establish that agents faithfully reproduce the full distribution of human responses. We will add a paragraph in §5.3 discussing this limitation explicitly, noting that the factorial design creates structured variation that both groups are expected to track, and that the more informative evidence for alignment comes from the coefficient comparison (Figure 5(b)) and the SES-stratified turnaround analysis (Figure 5(c)), which reveal both convergences and divergences. We will also strengthen the existing acknowledgment in the Limitations section regarding the small sample size. revision: yes

  3. Referee: §5.1, Figure 3(a): The reported correlation between network density and collective action success (r≈0.976) is likely inflated by shared time trends. Both network density and cooperation rate increase monotonically over the 25 rounds, so the high correlation may reflect co-evolution along a common trajectory rather than a meaningful structural relationship. The paper should consider differencing or detrending the time series before computing correlations, or at minimum acknowledge this concern.

    Authors: The referee raises a legitimate statistical concern. Both network density and cooperation rate increase monotonically over the 25 rounds, so the high Pearson correlation (r≈0.976) between them could be inflated by shared time trends. We will address this in two ways in the revision. First, we will compute first-difference correlations (correlating round-to-round changes in density with round-to-round changes in cooperation rate) and report these in the revised manuscript. If the differenced correlation remains substantial, this would support a genuine co-evolutionary relationship; if it drops substantially, this would confirm the referee's concern that the original correlation is largely driven by common trends. Second, regardless of the outcome, we will add an explicit caveat in §5.1 acknowledging that the high correlation may partly reflect shared temporal trajectories and that the more meaningful evidence for Putnam's theory comes from the co-evolution pattern itself—i.e., that network density and cooperation rise together in a self-reinforcing dynamic—rather than from the magnitude of the bivariate correlation. We will also revise the Observation text to frame the finding as co-evolution rather than a simple structural relationship. revision: yes

Circularity Check

1 steps flagged

Counterfactual 'causal lever' claim is partially built into the prompt architecture: trust is a direct numeric input to the adoption decision prompt, which explicitly instructs the agent to weigh trust before outputting ADOPT or REJECT.

specific steps
  1. fitted input called prediction [§6.2 (Counterfactual Intervention, Table 1) and Appendix F.3 (Technology Adoption Decision Prompt)]
    "§6.2: '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.' ... '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%.' ... 'suggesting that trust can serve as a key causal lever.' Appendix F.3 prompt: 'Platform Trust (how much you trust this technology provider): {platform_trust}/5.0' and 'General Trust in Community: {general_trust}/5.0' ... 'Belief: How trustworthy is the平台 and"

    The counterfactual intervention raises trust values, which are then fed directly as numeric inputs into the adoption decision prompt (Appendix F.3). The prompt explicitly instructs the LLM to reason about 'How trustworthy is the platform and the people recommending it?' as a 'Belief' step before outputting ADOPT or REJECT. Therefore, the direction of the effect (higher trust → higher adoption) is largely determined by the prompt's architecture: the LLM receives a higher trust number and is told to weigh it in its decision. The paper frames this as discovering that 'trust can serve as a key causal lever,' but the causal pathway is not an emergent property of the simulated social system — it is a designed input-output relationship in the prompt. The magnitude (15.4%) is not fully determined,

full rationale

The paper's counterfactual causal claim (§6.2) is partially circular. The intervention raises trust, which is a direct numeric input to the adoption decision prompt (Appendix F.3), and the prompt explicitly instructs the agent to reason about trust before deciding ADOPT or REJECT. The 'discovery' that trust increases adoption is thus close to a restatement of the prompt's design: the LLM is given a higher trust number and told to weigh it, so higher adoption is expected by construction. However, the specific magnitude (15.4%) and secondary effects (anxiety reduction, fewer decision contradictions) are not fully determined by the prompt architecture — they depend on the LLM's actual behavior. The macro-level pattern replication (§5.1) and human-agent alignment (§5.3) have more independent content: the alignment experiment validates that agents respond similarly to humans on the same stimuli, and the macro-level patterns emerge from multi-round interactions rather than a single prompt. No self-citation chain is load-bearing. The circularity is confined to the counterfactual intervention's causal framing, where the 'causal lever' is built into the decision function by prompt design.

Axiom & Free-Parameter Ledger

8 free parameters · 5 axioms · 3 invented entities

The framework introduces several hand-set parameters (N, T, τ, trust boost, norm decay threshold) without sensitivity analysis. The core assumption that LLM agents model human social behavior is the primary load-bearing axiom. The SST and SCM modules are new constructs whose validity depends entirely on the simulation's alignment with the 20-person human benchmark.

free parameters (8)
  • Number of agents (N) = 200 (applying task), 20 (modeling task)
    Chosen by the authors; not derived from theory.
  • Number of rounds (T) = 25
    Fixed simulation horizon; no sensitivity analysis provided.
  • Cooperation threshold (τ) = 0.5
    Default threshold for collective action success; stated as default without justification.
  • Temperature = 0.7
    LLM sampling temperature; affects behavioral variability.
  • top-p = 0.9
    LLM nucleus sampling parameter.
  • Trust boost in counterfactual = +1.0 (capped at 5.0)
    The intervention magnitude is chosen ad hoc for the counterfactual experiment.
  • Norm decay threshold = 10 consecutive rounds
    If no reciprocity observed over 10 rounds, norm decay triggers; this specific number is not justified.
  • Initial trust values = SES-correlated, e.g., low-SES ~2.0-2.6, high-SES ~3.1-4.3
    Hand-set initial values in Table 6; not derived from CGSS data directly.
axioms (5)
  • domain assumption LLM agents can faithfully model human social decision-making when given demographic and social-capital attributes.
    Invoked throughout §3.2 and §5.3; the entire framework depends on this being approximately true.
  • domain assumption Putnam's three dimensions (social network, trust, norms) are sufficient to capture collective-action dynamics.
    Stated in §2 and §3.1; the simulation only models these three, excluding other social factors.
  • standard math BDI (Belief-Desire-Intention) architecture is an appropriate formalism for LLM-based agent decision-making.
    Invoked in §3.2.2; adopted from Bratman (1987) as a background framework.
  • ad hoc to paper Text-based interaction is an adequate proxy for social interaction dynamics.
    Acknowledged as a limitation in the Limitations section; the simulation is entirely text-based.
  • domain assumption CGSS elderly subsample demographics are representative enough to initialize agent profiles.
    Used in §3.2.1 for SST module; the paper assumes this sample captures relevant socioeconomic stratification.
invented entities (3)
  • Social Structure Trait (SST) module no independent evidence
    purpose: Generates agents' structural social features from demographics
    A new module design; no external validation beyond the simulation itself.
  • Social Cognitive Memory (SCM) module no independent evidence
    purpose: Models adaptive learning and trust/norm updates
    New module; the update rules are defined by the authors and not independently validated against psychological literature on memory or learning.
  • Economic Connectivity (EC) index no independent evidence
    purpose: Quantifies cross-SES connections as proportion of high-SES friends in low-SES agent's network
    Introduced in §5.1; a new metric defined for this simulation without external benchmarking.

pith-pipeline@v1.1.0-glm · 37696 in / 3251 out tokens · 575339 ms · 2026-07-08T17:07:14.775547+00:00 · methodology

0 comments
read the original abstract

Putnam's Social Capital Theory is a foundational framework for collective action and community prosperity. However, traditional empirical methods face practical limits on control and replication. Meanwhile, LLM-based social simulations are typically behavior-driven and lack theory-aligned environments for modeling Putnam's core propositions. To address these gaps, we introduce SocaSim, an LLM-based multi-agent simulation framework to study Putnam's Social Capital Theory from theoretical blueprint to simulated reality. Specifically, we build an environment integrating social network evolution, trust dynamics, and norm propagation, where agents engage in repeated collective-action experiments, and then apply the three dimensions to analyze adaptation challenges in smart elderly care. Our simulations reproduce Putnam's macro-level patterns and exhibit strong human-agent alignment at the group level. Unlike traditional methods, SocaSim traces micro-level causal pathways of social network, trust, and norms via round-by-round simulations and counterfactual interventions, enabling process-level interpretability. Taken together, these capabilities establish a research paradigm that leverages LLM agents to bridge social science and computer science.

Figures

Figures reproduced from arXiv: 2607.06080 by Feng Ye, Hui Zheng, Shiyi Ling, Tong Xu, Wenjun Xue, Zhi Zheng.

Figure 1
Figure 1. Figure 1: Comparison of existing research paradigms [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of SOCASIM multi-agent architecture and round-by-round simulation workflow. framework to model, dynamically couple, and val￾idate the theory’s core elements. Specifically, we construct a social simulation system consisting of N agents, denoted as A = {a1, . . . , aN }, each equipped with demographic and socioeconomic attributes and initial social-capital endowments, sit￾uated in a controlled envir… view at source ↗
Figure 3
Figure 3. Figure 3: (a) demonstrates co-evolution between network density, cooperation rate, and collective action success. Over 25 rounds, network density in￾creased from 0.158 to 0.342, while cooperation rate (a) Social Networks and Collective Action (b) Trust Evolution across SES Groups (c) Norms and Cooperation Durability (d) Bonding and Bridging Social Capital [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Ablation study of SocaSim with SST, BDI, and SCM modules removed respectively. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Human-agent alignment: (a) scatter correlation, (b) coefficient comparison, (c) SES-switching heatmap. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Experimental simulation of social capital in smart elderly care technology adoption. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Case study: low-SES trust intervention. Metric Pre-intervention Post-intervention Relative change Technology Adoption Rate 0.732 ±0.028 0.845 ±0.024 +15.4% Psychological Pressure 0.268 ±0.032 0.215 ±0.028 -19.8% Anxiety Level 0.255 ±0.028 0.198 ±0.022 -22.4% Decision Contradictions 51 ±8 38 ±6 -25.5% [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Illustration and examples of three core dimensions in Putnam’s Social Capital Theory. [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: Prompt structure and the factorial design of social density, trust, and norms across eight scenarios. [PITH_FULL_IMAGE:figures/full_fig_p016_10.png] view at source ↗
Figure 10
Figure 10. Figure 10: Constructing Agent Profiles Through a Multi-Step Data Processing Pipeline of CGSS 2023 Dataset. [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗
Figure 13
Figure 13. Figure 13: Demographic distribution of the simulated elderly population. [PITH_FULL_IMAGE:figures/full_fig_p018_13.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

42 extracted references · 42 canonical work pages · 5 internal anchors

  1. [1]

    Putnam and Robert Leonardi and Raffaella Y

    Robert D. Putnam and Robert Leonardi and Raffaella Y. Nonetti. Making Democracy Work: Civic Traditions in Modern Italy. 1994

    work page 1994

  2. [2]

    Robert D. Putnam. Bowling Alone: The Collapse and Revival of American Community. 2000

    work page 2000

  3. [3]

    Social capital and education -- An attempt to synthesize conceptualization arising from various theoretical origins

    Piotr Mikiewicz. Social capital and education -- An attempt to synthesize conceptualization arising from various theoretical origins. Cogent Education. 2021. doi:10.1080/2331186X.2021.1907956

    work page doi:10.1080/2331186x.2021.1907956 2021

  4. [4]

    The outcomes of social capital among event runners: quality of life considerations

    Ran Zhou and Kyriaki Kaplanidou. The outcomes of social capital among event runners: quality of life considerations. International Journal of Sports Marketing and Sponsorship. 2023. doi:10.1108/IJSMS-03-2023-0045

    work page doi:10.1108/ijsms-03-2023-0045 2023

  5. [5]

    The Impact of Social Capital on Workplace Engagement, Productivity, and Organizational Success

    Do Hong Thang. The Impact of Social Capital on Workplace Engagement, Productivity, and Organizational Success. International Journal of Humanities Social Sciences and Education. 2025. doi:10.20431/2349-0381.1202004

    work page doi:10.20431/2349-0381.1202004 2025

  6. [6]

    Unpacking Social Capital

    Ruben Durante and Nicola Mastrorocco and Luigi Minale and James M Snyder. Unpacking Social Capital. The Economic Journal. 2025. doi:10.1093/ej/ueae074

    work page doi:10.1093/ej/ueae074 2025

  7. [7]

    Indirect impact of employees' social capital on their income at Dong Thap Muoi, Long An Province

    Pham Tan Hoa. Indirect impact of employees' social capital on their income at Dong Thap Muoi, Long An Province. HCMCOUJS-Economics and Business Administration. 2021. doi:10.46223/HCMCOUJS.econ.en.11.2.1952.2021

    work page doi:10.46223/hcmcoujs.econ.en.11.2.1952.2021 2021

  8. [8]

    Senathirajah

    Ayesha Akter Sumi and Jaffary Bin Awang and Rasheedul Haque and Sayeeduzzafar Qazi and Nur Farhana Abdul Rahman and Abdul Rahman Bin S. Senathirajah. Investigating the function of religion and social capital in shaping sustainable social development. Discover Sustainability. 2025. doi:10.1007/s43621-025-01622-x

    work page doi:10.1007/s43621-025-01622-x 2025

  9. [9]

    Agents of change---together: Using agent-based models to inspire social capital building for resilient communities

    Linda Shenk and Caroline Krejci and Ulrike Passe. Agents of change---together: Using agent-based models to inspire social capital building for resilient communities. Community Development. 2019. doi:10.1080/15575330.2019.1574849

    work page doi:10.1080/15575330.2019.1574849 2019

  10. [10]

    The Concept and Application of Social Capital in Health, Education and Employment: A Scoping Review

    Syden Mishi and Kin Sibanda and Godfred Anakpo. The Concept and Application of Social Capital in Health, Education and Employment: A Scoping Review. Social Sciences. 2023. doi:10.3390/socsci12080450

    work page doi:10.3390/socsci12080450 2023

  11. [11]

    The Role of Social Capital in Agricultural and Rural Development: Lessons Learnt from Case Studies in Seven Countries

    Maria Rivera and Karlheinz Knickel and José María Díaz-Puente and Ana Afonso. The Role of Social Capital in Agricultural and Rural Development: Lessons Learnt from Case Studies in Seven Countries. Sociologia Ruralis. 2019. doi:10.1111/soru.12218

    work page doi:10.1111/soru.12218 2019

  12. [12]

    LLM -Based Edge Intelligence: A Comprehensive Survey on Architectures, Applications, Security and Trustworthiness

    Othmane Friha and Mohamed Amine Ferrag and Burak Kantarci and Burak Cakmak and Arda Ozgun and Nassira Ghoualmi-Zine. LLM -Based Edge Intelligence: A Comprehensive Survey on Architectures, Applications, Security and Trustworthiness. IEEE Open Journal of the Communications Society. 2024. doi:10.1109/OJCOMS.2024.3456549

    work page doi:10.1109/ojcoms.2024.3456549 2024

  13. [13]

    User Behavior Simulation with Large Language Model-based Agents

    Lei Wang and Jingsen Zhang and Hao Yang and Zhi-Yuan Chen and Jiakai Tang and Zeyu Zhang and Xu Chen and Yankai Lin and Hao Sun and Ruihua Song and Xin Zhao and Jun Xu and Zhicheng Dou and Jun Wang and Ji-Rong Wen. User Behavior Simulation with Large Language Model-based Agents. ACM Trans. Inf. Syst. 2025. doi:10.1145/3708985

    work page doi:10.1145/3708985 2025

  14. [14]

    AgentSociety: Large-Scale Simulation of LLM-Driven Generative Agents Advances Understanding of Human Behaviors and Society

    Jinghua Piao and Yuwei Yan and Jun Zhang and Nian Li and Junbo Yan and Xiaochong Lan and Zhihong Lu and Zhiheng Zheng and Jing Yi Wang and Di Zhou and Chen Gao and Fengli Xu and Fang Zhang and Ke Rong and Jun Su and Yong Li , year=. 2502.08691 , archivePrefix=

    work page internal anchor Pith review Pith/arXiv arXiv

  15. [15]

    OASIS: Open Agent Social Interaction Simulations with One Million Agents

    Ziyi Yang and Zaibin Zhang and Zirui Zheng and Yuxian Jiang and Ziyue Gan and Zhiyu Wang and Zijian Ling and Jinsong Chen and Martz Ma and Bowen Dong and Prateek Gupta and Shuyue Hu and Zhenfei Yin and Guohao Li and Xu Jia and Lijun Wang and Bernard Ghanem and Huchuan Lu and Chaochao Lu and Wanli Ouyang and Yu Qiao and Philip Torr and Jing Shao , year=. 2...

    work page internal anchor Pith review Pith/arXiv arXiv

  16. [16]

    CompeteAI : Understanding the Competition Dynamics of Large Language Model-Based Agents

    Qinlin Zhao and Jindong Wang and Yixuan Zhang and Yiqiao Jin and Kaijie Zhu and Hao Chen and Xing Xie. CompeteAI : Understanding the Competition Dynamics of Large Language Model-Based Agents. Proceedings of the 41st International Conference on Machine Learning (ICML 2024). 2024

    work page 2024

  17. [17]

    McNamara and Deming Chen , year=

    Jingru Jia and Zehua Yuan and Junhao Pan and Paul E. McNamara and Deming Chen , year=. 2502.20432 , archivePrefix=

    work page arXiv

  18. [18]

    SOTOPIA - : Dynamic Strategy Injection Learning and Social Instruction Following Evaluation for Social Agents

    Wenyuan Zhang and Tianyun Liu and Mengxiao Song and Xiaodong Li and Tingwen Liu. SOTOPIA - : Dynamic Strategy Injection Learning and Social Instruction Following Evaluation for Social Agents. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025. doi:10.18653/v1/2025.acl-long.1203

    work page doi:10.18653/v1/2025.acl-long.1203 2025

  19. [19]

    Social Capital in Bourdieu’s, Coleman’s and Putnam’s Theory: Empirical Evidence and Emergent Measurement Issues

    Michael Tzanakis. Social Capital in Bourdieu’s, Coleman’s and Putnam’s Theory: Empirical Evidence and Emergent Measurement Issues. Educate. 2013

    work page 2013

  20. [20]

    Dodd and John Brummette and Vincent Hazleton

    Melissa D. Dodd and John Brummette and Vincent Hazleton. A social capital approach: An examination of Putnam's civic engagement and public relations roles. Public Relations Review. 2015. doi:10.1016/j.pubrev.2015.05.001

    work page doi:10.1016/j.pubrev.2015.05.001 2015

  21. [21]

    The role of open innovation and a normalizing mechanism of social capital in the tourism industry

    Sanmugam Annamalah and Pradeep Paraman and Selim Ahmed and Rajinder Dass and Ilham Sentosa and Thillai Raja Pertheban and Farid Shamsudin and Baharudin Kadir and Kalisri Logeswaran Aravindan and Murali Raman and Wong Chee Hoo and Prakash Singh. The role of open innovation and a normalizing mechanism of social capital in the tourism industry. Journal of Op...

    work page doi:10.1016/j.joitmc.2023.100056 2023

  22. [22]

    Carpiano

    Richard M. Carpiano. Toward a neighborhood resource-based theory of social capital for health: Can Bourdieu and sociology help?. Social Science & Medicine. 2006. doi:10.1016/j.socscimed.2005.05.020

    work page doi:10.1016/j.socscimed.2005.05.020 2006

  23. [23]

    Roger D. Goddard. Relational Networks, Social Trust, and Norms: A Social Capital Perspective on Students' Chances of Academic Success. Educational Evaluation and Policy Analysis. 2003. doi:10.3102/01623737025001059

    work page doi:10.3102/01623737025001059 2003

  24. [24]

    Instrument Development for Socio-Educational Support and Socio-Educational Participation towards Educational Well-Being

    Jalal Careemdeen and Mohd Awang and Abdul Ahmad and Shakila Dahalan. Instrument Development for Socio-Educational Support and Socio-Educational Participation towards Educational Well-Being. Journal of Educational and Social Research. 2021. doi:10.36941/jesr-2021-0104

    work page doi:10.36941/jesr-2021-0104 2021

  25. [25]

    Bonding, Bridging and Linking Capital: How Social Capital was Mobilized During the PPKM Micro in Malang Raya

    Anif Chawa and Moch Putra and Anik Susanti. Bonding, Bridging and Linking Capital: How Social Capital was Mobilized During the PPKM Micro in Malang Raya. Jurnal Ilmu Sosial. 2024. doi:10.14710/jis.22.2.2023.67-88

    work page doi:10.14710/jis.22.2.2023.67-88 2024

  26. [26]

    2502.09082 , archivePrefix=

    Xintao Wang and Heng Wang and Yifei Zhang and Xinfeng Yuan and Rui Xu and Jen-tse Huang and Siyu Yuan and Haoran Guo and Jiangjie Chen and Shuchang Zhou and Wei Wang and Yanghua Xiao , year=. 2502.09082 , archivePrefix=

    work page arXiv

  27. [27]

    Beyond Profile: From Surface-Level Facts to Deep Persona Simulation in LLM s

    Zixiao Wang and Duzhen Zhang and Ishita Agarwal and Shen Gao and Le Song and Xiuying Chen. Beyond Profile: From Surface-Level Facts to Deep Persona Simulation in LLM s. Findings of the Association for Computational Linguistics: ACL 2025. 2025. doi:10.18653/v1/2025.findings-acl.1094

    work page doi:10.18653/v1/2025.findings-acl.1094 2025

  28. [28]

    Sudhakar and Hadi Nekoei and Mathieu Reymond and Miao Liu and Janarthanan Rajendran and Sarath Chandar

    Arjun V. Sudhakar and Hadi Nekoei and Mathieu Reymond and Miao Liu and Janarthanan Rajendran and Sarath Chandar. A Generalist Hanabi Agent. International Conference on Learning Representations. 2025

    work page 2025

  29. [29]

    Findings of the Association for Computational Linguistics: NAACL 2025

    Zeyu Zhang and Jianxun Lian and Chen Ma and Yaning Qu and Ye Luo and Lei Wang and Rui Li and Xu Chen and Yankai Lin and Le Wu and Xing Xie and Ji-Rong Wen , title =. Findings of the Association for Computational Linguistics: NAACL 2025. 2025. doi:10.18653/v1/2025.findings-naacl.160

    work page doi:10.18653/v1/2025.findings-naacl.160 2025

  30. [30]

    MANTRA: Enhancing Automated Method-Level Refactoring with Contextual RAG and Multi-Agent LLM Collaboration

    Yisen Xu and Feng Lin and Jinqiu Yang and Tse-Hsun and Chen and Nikolaos Tsantalis , year=. 2503.14340 , archivePrefix=

    work page internal anchor Pith review Pith/arXiv arXiv

  31. [31]

    Bratman , title =

    Michael E. Bratman , title =. 1987

    work page 1987

  32. [32]

    Collective Action and the Evolution of Social Norms

    Elinor Ostrom. Collective Action and the Evolution of Social Norms. Journal of Economic Perspectives. 2000. doi:10.1257/jep.14.3.137

    work page doi:10.1257/jep.14.3.137 2000

  33. [33]

    Trusting the State, Trusting Each Other? The Effect of Institutional Trust on Social Trust

    Kim Mannemar Sønderskov and Peter Thisted Dinesen. Trusting the State, Trusting Each Other? The Effect of Institutional Trust on Social Trust. Political Behavior. 2016. doi:10.1007/s11109-015-9322-8

    work page doi:10.1007/s11109-015-9322-8 2016

  34. [34]

    2025 , eprint=

    Qwen2.5 Technical Report , author=. 2025 , eprint=

    work page 2025

  35. [35]

    OpenAI and Josh Achiam and Steven Adler and Sandhini Agarwal and Lama Ahmad and Ilge Akkaya and Florencia Leoni Aleman and Diogo Almeida and Janko Altenschmidt and Sam Altman and Shyamal Anadkat and Red Avila and Igor Babuschkin and Suchir Balaji and Valerie Balcom and Paul Baltescu and Haiming Bao and Mohammad Bavarian and Jeff Belgum and Irwan Bello and...

    work page internal anchor Pith review Pith/arXiv arXiv

  36. [36]

    ChatGLM: A Family of Large Language Models from GLM-130B to GLM-4 All Tools

    Team GLM: Aohan Zeng and Bin Xu and Bowen Wang and Chenhui Zhang and Da Yin and Dan Zhang and Diego Rojas and Guanyu Feng and Hanlin Zhao and Hanyu Lai and Hao Yu and Hongning Wang and Jiadai Sun and Jiajie Zhang and Jiale Cheng and Jiayi Gui and Jie Tang and Jing Zhang and Jingyu Sun and Juanzi Li and Lei Zhao and Lindong Wu and Lucen Zhong and Mingdao L...

    work page internal anchor Pith review Pith/arXiv arXiv

  37. [37]

    Can Large Language Model Agents Simulate Human Trust Behavior?

    Chengxing Xie and Canyu Chen and Feiran Jia and Ziyu Ye and Shiyang Lai and Kai Shu and Jindong Gu and Adel Bibi and Ziniu Hu and David Jurgens and James Evans and Philip Torr and Bernard Ghanem and Guohao Li. Can Large Language Model Agents Simulate Human Trust Behavior?. Advances in Neural Information Processing Systems. 2024. doi:10.52202/079017-0501

    work page doi:10.52202/079017-0501 2024

  38. [38]

    Investigating Prosocial Behavior Theory in LLM Agents Under Policy-Induced Inequities

    Yujia Zhou and Hexi Wang and Qingyao Ai and Zhen Wu and Yiqun Liu. Investigating Prosocial Behavior Theory in LLM Agents Under Policy-Induced Inequities. Proceedings of the Fortieth AAAI Conference on Artificial Intelligence (AAAI 2026). 2026

    work page 2026

  39. [39]

    Age Sustainability in Smart City: Seniors as Urban Stakeholders in the Light of Literature Studies

    Izabela Jonek-Kowalska and Maciej Wolny. Age Sustainability in Smart City: Seniors as Urban Stakeholders in the Light of Literature Studies. Sustainability. 2025. doi:10.3390/su17146333

    work page doi:10.3390/su17146333 2025

  40. [40]

    Analysis of China's Smart Elderly Care Service Policy: Based on the Three-dimensional Framework

    Xiaokun Sun. Analysis of China's Smart Elderly Care Service Policy: Based on the Three-dimensional Framework. SAGE Open. 2024. doi:10.1177/21582440241240239

    work page doi:10.1177/21582440241240239 2024

  41. [41]

    Gustavson

    Mohammad Mahdi Fakhimi and Adriana Hughes and Allison M. Gustavson. Evaluating Smart Home Usability and Accessibility in Early Detection and Intervention of Mental Health Challenges Among Older Adults: A Narrative Review and Framework. Journal of Ageing and Longevity. 2025. doi:10.3390/jal5010003

    work page doi:10.3390/jal5010003 2025

  42. [42]

    Boyd and Zhijing Jin and Veronica Perez-Rosas and Steven Wilson and James W

    Rada Mihalcea and Laura Biester and Ryan L. Boyd and Zhijing Jin and Veronica Perez-Rosas and Steven Wilson and James W. Pennebaker. How developments in natural language processing help us in understanding human behaviour. Nature Human Behaviour. 2024. doi:10.1038/s41562-024-01938-0

    work page doi:10.1038/s41562-024-01938-0 2024