[论文解读] Do LLM-Driven Agents Exhibit Engagement Mechanisms? Controlled Tests of Information Load, Descriptive Norms, and Popularity Cues

Tai-Quan Peng, Yuan Tian+3 more|arXiv (Cornell University)|Mar 21, 2026

Language and cultural evolution被引用 0

一句话总结

This study uses an LLM-driven Weibo-like simulation to test whether information load, descriptive norms, and endogenous popularity cues produce theoretically grounded engagement patterns, not just plausible behavior. It finds load-sensitive engagement and bandwagon effects that vary by normative context.

ABSTRACT

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.

研究动机与目标

• Motivate the use of LLM-driven simulations to test engagement mechanisms in social media contexts.
• Examine whether information load, descriptive norms, and popularity cues produce theoretically interpretable engagement patterns rather than mere plausible traces.
• Disentangle participation thresholds from engagement allocation to identify mechanism-specific effects.
• Assess how endogenous popularity signals (bandwagon cues) interact with experimental manipulations across contexts.

提出的方法

• Develop an LLM-powered agent-based simulation of a Sina Weibo–like environment with 558 agents and seed posts.
• Manipulate information load by varying the number of posts shown per activation while keeping network content constant.
• Manipulate descriptive norms via prompts that frame the prevalence of engagement forms (like vs repost); include a no-norm baseline.
• Allow popularity cues (cumulative likes and reshares) to evolve endogenously and influence decisions.
• Treat engagement as a two-stage outcome: participation threshold (engage vs read) and engagement allocation (like, repost, quote).
• Analyze using two-stage regression: binary logistic for engagement vs reading, and multinomial logistic for form choice.

Figure 1: Observed distribution of engagement behaviors across normative regimes and information load levels.

实验结果

研究问题

• RQ1 RQ1: Do LLM-driven agents show load-sensitive changes in engagement probability consistent with capacity-based expectations?
• RQ2 RQ2: Do popularity cues increase engagement likelihood and amplify response probabilities consistent with bandwagon effects?
• RQ3 RQ3: Do descriptive norms reorganize engagement allocation as signaled by normative prompts?
• RQ4 RQ4: How do information load, descriptive norms, and popularity cues jointly shape engagement thresholds and allocations?

主要发现

• Engagement likelihood is strongly associated with popularity cues and shows a negative gradient with higher information load at baseline conditions.
• Descriptive norm prompts alter engagement allocation, producing different hierarchies of engagement forms across norm regimes.
• A significant three-way interaction among popularity cues, information load, and repost-dominant norms reveals conditional bandwagon effects where high popularity cues can overcome high information load to boost engagement.
• No-norm and like-dominant regimes share a load-sensitive pattern, whereas repost-dominant norms show stronger responsiveness to popularity cues under higher load.
• Engagement allocation shows distinct patterns: like-dominant norms favor liking, repost-dominant norms favor reposting, and the effects interact with load and popularity cues.

Figure 2: Predicted probability of engagement (vs. read-only) adjusted by information load, popularity cues, and normative regimes.