Organizations: Department of Communication, Michigan State University, USA · College of Computer Science and Technology, Zhejiang University, China · School of Software Technology, Zhejiang University, China
Large language models make agent-based simulation more behaviorally expressive, but they also sharpen a basic methodological tension: fluent, human-like output is not, by itself, evidence for theory. We evaluate what an LLM-driven simulation can credibly support using information engagement on social media as a test case. In a Weibo-like environment, we manipulate information load and descriptive norms, while allowing popularity cues (cumulative likes and Sina Weibo-style cumulative reshares) to evolve endogenously. We then ask whether simulated behavior changes in theoretically interpretable ways under these controlled variations, rather than merely producing plausible-looking traces. Engagement responds systematically to information load and descriptive norms, and sensitivity to popularity cues varies across contexts, indicating conditionality rather than rigid prompt compliance. We discuss methodological implications for simulation-based communication research, including multi-condition stress tests, explicit no-norm baselines because default prompts are not blank controls, and design choices that preserve endogenous feedback loops when studying bandwagon dynamics.
Figures & tables
Study Component
Origin
Role in Design
Agent profiles ( n=558 )
Real-world data (Sina Weibo)
Provide topic-specific user attributes and network structure
Follower–following relationships
Real-world data (partial network)
Establish diffusion pathways and content visibility channels
Seed posts ( n=50 , Chinese)
Real-world data (DeepSeek discussions)
Initialize cascades with empirically grounded content
Information load (4 conditions)
Experimentally manipulated
Vary choice-set size at each activation (posts shown)
Descriptive norm regime (3 conditions)
Experimentally manipulated
Manipulate perceived engagement prevalence via prompt framing
Agent activation probability (1% per timestep)
Controlled simulation parameter
Hold activity rate constant across conditions
Table 1: Summary of Empirical Inputs, Experimental Manipulations, and Simulated Components
Figure 1: Observed distribution of engagement behaviors across normative regimes and information load levels.
Predictor
B
SE
Odds Ratio
p -value
Intercept
-2.787
0.041
0.062
<.001
Popularity Cues a
4.624
0.115
101.871
<.001
Low Load (vs. Lowest Load) b
-0.438
0.056
0.645
<.001
Medium Load (vs. Lowest Load) b
-0.758
0.051
0.469
<.001
High Load (vs. Lowest Load) b
-1.297
0.049
0.273
<.001
Like-dominant Norm (vs. No Norm) c
0.291
0.057
1.338
<.001
Table 2: Binary logistic regression of engagement on prior cumulative engagement, information load, and descriptive norm regime
Figure 2: Predicted probability of engagement (vs. read-only) adjusted by information load, popularity cues, and normative regimes.
Quote (vs. Like)
Repost (vs. Like)
Predictor
B
SE
OR
p
B
SE
OR
p
Intercept
-0.180
0.081
0.835
.027
-4.108
0.380
0.016
<.001
Popularity cues a
-0.935
0.098
0.392
<.001
0.141
0.338
1.152
.676
Low Load (vs. Lowest Load) b
-0.418
0.112
0.658
<.001
-0.084
0.498
0.919
.865
Medium Load (vs. Lowest Load) b
-1.522
0.122
0.218
<.001
0.262
0.537
1.300
.626
High Load (vs. Lowest Load) b
-1.683
0.116
0.186
<.001
-0.595
0.528
0.552
.259
Table 3: Multinomial logistic regression of engagement allocation on popularity cues, information load, and descriptive norm regime.
Figure 3: Predicted probabilities of liking, reposting, and quoting across normative regimes and information load conditions.
Figure 4: Predicted probabilities of liking, reposting, and quoting across normative regimes and popularity cues.
RQ
Mechanism focus
Engagement stage
Core finding
Implication
RQ1
Information load
Threshold (engage vs. read)
Higher information load generally reduces the likelihood of transitioning from reading to any engagement, especially in the no-norm and like-dominant regimes.
Participation thresholds are load-sensitive.
RQ2
Popularity cues / bandwagon effects
Threshold (primary) and allocation (secondary)
Stronger popularity cues substantially increase the likelihood of engagement, consistent with bandwagon effects; conditional on engagement, popularity cues are associated with greater concentration on focal actions within regimes.
Popularity cues shape both participation and allocation.
RQ3
Descriptive norms
Allocation (like/repost/quote)
Descriptive norm prompts systematically reorganize engagement allocation toward the normatively emphasized action without enforcing fixed proportions.
Norms reorganize engagement composition.
RQ4
Joint effects
Threshold and allocation
Information load, descriptive norms, and popularity cues interact: the association between popularity cues and engagement varies across norm regimes and load levels, with the repost-dominant regime showing crossover patterns in which higher load can amplify responsiveness to popularity cues.
Mechanisms are conditional, not additive.
Table 4: Summary of Core Findings Across Research Questions
Thomas Lord Department of Computer Science, University of Southern California · Annenberg School for Communication and Journalism, University of Southern California · Marshall School of Business, University of Southern California