Many researchers now build social simulations from large language model (LLM) agents: each simulated user is a chatbot with a persona, placed in a follower network, deciding what to repost. Real human sharing follows a well-measured pattern. On a large microblogging network, the chance that someone forwards a message peaks after two exposures and then falls when the exposures all come from one tight-knit group. Very large cascades are instead driven by exposures from unconnected sources in different communities (Cumulative Effect in Information Diffusion). LLM agents appear to behave the opposite way. In repeated conversation they drift toward agreement and flattery, and stronger models pull weaker ones toward their style in one direction only (Attractor States Emerge in Multi-Turn LLM Conversations). Their personas also collapse toward a generic average (The Persona Fidelity Gap). If simulated users become more willing to repost with every repeated exposure, and all respond alike, then a simulator will get wrong the very thing it is built to study: whether information stays inside one community or crosses into many.
We propose a replay test and a targeted repair. For the test, we take real users from the Sina Weibo corpus, infer each user's persona from their own past posts, and feed the matching LLM agent the exact sequence of exposures that user actually received. We then compare the agent's repost decisions with what the user did, broken down by how many exposures arrived and how many distinct communities they came from. For the repair, we add a learned, directed receptivity model. It estimates how much a sender with one inferred persona moves a receiver with another, and how quickly that influence wears off when further exposures come from the receiver's already-seen community. The claim is specific. Influence that is directed (sender-to-receiver weight differs from the reverse), conditioned on inferred persona, and sensitive to community structure is the ingredient that lets LLM simulators reproduce human exposure-response curves. Neither a better persona prompt nor a global structural-diversity rule supplies it.
Core Claim
Structural diversity is the idea that exposures from distinct, disconnected communities matter more than repeated exposures from one group. So far it has been measured as pure topology, with every receiver assumed equally susceptible (Toward Order-of-Magnitude Cascade Prediction; An Empirical Evaluation Of Social Influence Metrics). Implicit persona inference recovers a user's hidden traits and preferences from behavior instead of from a written profile. It has been developed only for one-on-one dialogue (Simulating Before Planning; Learning to Predict Persona Information). Asymmetric influence modeling separates the influencer from the influenced. It has mostly been studied in pairs or small groups (Content-based Modeling of Reciprocal Relationships; Yue Hu et al.).
We combine the three as follows. Structural diversity is the behavioral target a simulator must reproduce. Persona inference supplies receiver-specific susceptibility. Asymmetry is the form the influence weights must take. Concretely, the probability that user $u$ reposts message $m$ after exposures from senders $v_1,\dots,v_K$ is modeled as
$$ P(\text{repost}) = \sigma\Big( s_{\text{LLM}}(z_u, m) + b(z_u) + \sum_{k=1}^{K} w(z_{v_k} \to z_u)\, \rho(z_u)^{n_k} \Big). $$
The symbols are as follows. $\sigma$ is the logistic function. $z_u$ and $z_{v_k}$ are persona embeddings inferred from each user's posting history. $s_{\text{LLM}}$ is the agent's judgment of how well the message content fits the persona. $b(z_u)$ is a baseline sharing propensity. $w(\cdot \to \cdot)$ is a directed influence weight, with $w(z_v \to z_u) \neq w(z_u \to z_v)$ in general. $n_k$ counts earlier exposures that came from sender $v_k$'s community. $\rho(z_u) \in (0,1)$ is a persona-dependent discount on redundant same-community exposures. The two-exposure peak and the advantage of diverse motifs both follow from $\rho < 1$. The proposal's bet is that $\rho$ and $w$ vary systematically with inferred persona. If they do not, a global constant should do as well, and the experiments include that comparison.
Relation to Prior Work
The closest simulation work is the seed paper LLM-Augmented Agent-Based Simulation of Cyber Social Agents. It places empirically parameterized archetypes of automated accounts into an LLM network simulation to study how agent diversity shapes diffusion. Its personas are fixed archetypes, and its outputs are not checked against real exposure-response curves. We infer personas per user and make curve fidelity the primary metric. Population-Aligned Persona Generation motivates this work and supplies a baseline. It matches synthetic persona pools to real Big Five distributions, cutting alignment error by 49.8%, and names network diffusion as future work. It aligns the population's marginal distribution but not individual users. Our replay requires the latter.
Two papers motivate the hypothesized failure. The Persona Fidelity Gap shows that even history-grounded LLM personas cover only 4.6–9.4% of real users' preference neighborhoods. Within groups, LLM personas reach cosine similarities of 0.94–0.99, compared with 0.56–0.68 for humans. Attractor States shows affiliation drift in debates between two LLMs. Mixed-model play cuts endpoint distance by 23.6%, and one model dominates another (a dominance score of −0.538 for Claude Haiku against Nemotron). Neither paper places agents on a network. Together they predict the tension we test: monotonic capitulation in LLMs against saturation in humans.
Three lines of work supply the method. Persona encoders that map raw history into persona space come from Wangchunshu Zhou et al. and from the history denoiser in UDP (Simulating Before Planning). The split between sender significance and receiver receptivity comes from the Hawkes–Gaussian-process model (Content-based Modeling of Reciprocal Relationships). That model relies on unigram divergences and Markov chain Monte Carlo sampling, so we replace both with learned embeddings trained by gradient descent. Ordered concatenation of influencer and influenced features follows Yue Hu et al.. We use the seed paper PB-GRPO only in an optional variant: it normalizes reinforcement-learning advantages within buckets of similar users, and we borrow that bucketing to fine-tune the agent's own repost judgment on real outcomes. PB-GRPO itself has no network.
ComPO conditions a language model on a known community label. We infer persona instead of reading off a label, and we model influence between individuals, not community norms.
Research Program
The work has three stages.
The first stage builds the replay benchmark on the Weibo corpus. The corpus contains 16.6M original messages, 58.6M users and 265.5M directed follow links (Cumulative Effect in Information Diffusion). For each sampled user and message, we reconstruct the time-ordered exposures from followees who reposted before the user's decision. Communities come from Louvain partitions, as in Toward Order-of-Magnitude Cascade Prediction, with each exposure tagged by its source community. The target is the empirical adoption surface $P(\text{repost} \mid K \text{ exposures}, C \text{ distinct communities})$, together with the second-source echo rate, which the source paper reports as 85.5% in two-exposure cases. Users are split by person, so that test users' histories never touch training.
The second stage diagnoses existing simulators. It runs replay with agents built in four ways: no persona, demographic or archetype prompts in the style of the Cyber Social Agents work, population-aligned personas, and history-grounded persona prompts. Each agent receives the exposures as successive feed items. We measure where each simulated adoption surface departs from the human one. We also measure how much receivers differ from one another, using the coverage metric from The Persona Fidelity Gap applied to per-user adoption profiles.
The third stage trains the receptivity model on training users' real repost logs and plugs it into the agent's decision. The LLM still writes the content and supplies the content-fit term. Ablations remove persona conditioning, remove the directed ordering, or remove the community discount. Finally, we run forward simulations, seeding real early cascades, to test whether the corrected agents produce realistic cascade-level outcomes.
Baselines
The naive reference point is a logistic model on exposure count alone, plus the global base repost rate. It encodes the cumulative-effect assumption. Any method that cannot beat it on the adoption surface has learned nothing about structure. It is fit on the same training users.
The strongest existing method for predicting adoption and cascades is the feature-based Random Forest. It uses Louvain structural-diversity features of adopters, of the recently exposed frontier, and of non-adopters. On Weibo, it reaches precision 0.69 and recall 0.52 for predicting tenfold cascade growth (Toward Order-of-Magnitude Cascade Prediction). A multi-feature Random Forest for influence reached precision 0.95 and F1 0.948, against 0.679 and 0.622 for a neighbor-count baseline (An Empirical Evaluation Of Social Influence Metrics). It is the strongest competitor because it already captures the structural-diversity effect without any persona or LLM. We retrain it on our exact splits and Louvain partitions, with the same hyperparameter search budget as our receptivity model. We also give it the same user history in the form of a bag-of-words or sentence-embedding vector, so that any gain we report is not merely access to text. Beating it means higher held-out AUC for individual adoption on cold-start users, meaning test users unseen in training, and equal or better precision and recall on the tenfold-growth task.
The strongest LLM-simulation baselines are archetype personas in the style of the Cyber Social Agents work, and population-aligned personas from Population-Aligned Persona Generation. Both use the same LLM, prompts, exposure formatting and inference budget as our agents. No reported result on Lacuna exists for either on Weibo replay or exposure-response fidelity. We will report the first numbers. Beating them means a lower divergence between simulated and real adoption surfaces, and higher coverage of receiver heterogeneity.
The key ablation replaces $w(z_v \to z_u)$ and $\rho(z_u)$ with persona-independent constants fit on the same data. It is the most important baseline because it isolates the proposal's central ingredient. If it matches the full model, persona-conditioned asymmetry adds nothing beyond a global structural-diversity rule. A second ablation symmetrizes $w$, so that each pair gets one shared weight.
Why not just…?
Why not just use the feature-based Random Forest and drop the LLM entirely? For predicting who reposts, this is a real threat, and it may win on warm users whose identities appear in training. The proposal is not mainly a better predictor. It aims to make simulators faithful, and simulators need agents that write and rewrite content, respond to unseen messages, and generalize to users without logs. The cold-start split settles the question. If the Random Forest, given the same history embeddings, matches our model on new users and new messages, we will report that LLM agents add nothing for adoption and restrict our claim to content generation.
Why not just give agents a better persona prompt, grounded in history or population-aligned? History grounding helps, cutting the fidelity gap 30-fold relative to demographic templates. Yet it still leaves coverage under 10% and within-group similarity near 0.94–0.99 (The Persona Fidelity Gap). Population alignment fixes marginal distributions, not individual responses (Population-Aligned Persona Generation). Neither addresses drift toward agreement under repeated exposure. Both are run as baselines in stage two, and the adoption surface will show whether either already yields the two-exposure peak.
Why not just tell the agent the structure, for example "you have already seen this from three people in your circle," or prompt it to discount redundancy? This could plausibly correct the average curve shape, and we test it as a prompt-only baseline. Our prediction is that it produces one uniform discount for everyone, so it should match the persona-free ablation and fail on receiver heterogeneity. If it matches the full model, the learned receptivity component is unnecessary, and we would recommend the prompt.
Why not just use symmetric influence weights? Follow links are directed, and both dyadic LLM evidence (Attractor States) and reciprocity modeling (Content-based Modeling of Reciprocal Relationships) show strong directionality. Whether this matters for cascade adoption is untested. The symmetric ablation measures it directly.
Decisive Checks
The first check is whether off-the-shelf LLM agents show a monotonically rising or flat adoption curve as same-community exposures accumulate, where humans show a peak at two. If they already reproduce the human peak, the diagnostic motivation fails, and we report that as a positive finding about current simulators. The second check is whether the full model beats the persona-independent ablation on held-out per-user adoption AUC and on surface divergence, specifically among receivers in the tails of the inferred persona distribution. That is where persona collapse predicts the largest errors. The third check is whether forward simulations seeded with real early cascades reproduce the dominance of diverse motifs in large cascades, and reach precision and recall on tenfold growth at least comparable to 0.69 and 0.52. Failing the second check while passing the first would mean a structure-aware prompt or constant discount suffices. We would then publish the replay benchmark and the negative result on persona conditioning.
Workshop authors to connect with
- LLM-Augmented Agent-Based Simulation of Cyber Social Agents: Lynnette Hui Xian Ng, Kathleen M. Carley
- PB-GRPO: Learning Socially Adaptive LLM Agents from Persona-Driven Simulation with Preference-Batched GRPO: Jingquan Wang, Jun Yin, Xu Han, Yongsheng Mei, Jie Hao, Bin Guo
References
- Cumulative Effect in Information Diffusion: Empirical Study on a Microblogging Network
- Attractor States Emerge in Multi-Turn LLM Conversations
- The Persona Fidelity Gap: Behaviorally Grounded LLM Personas Still Compress Real-User Preference Diversity
- Toward Order-of-Magnitude Cascade Prediction
- An empirical evaluation of social influence metrics
- Simulating Before Planning: Constructing Intrinsic User World Model for User-Tailored Dialogue Policy Planning
- Learning to Predict Persona Information for Dialogue Personalization without Explicit Persona Description
- Content-based Modeling of Reciprocal Relationships using Hawkes and Gaussian Processes
- Collaborative Motion Prediction via Neural Motion Message Passing
- LLM-Augmented Agent-Based Simulation of Cyber Social Agents
- Population-Aligned Persona Generation for LLM-based Social Simulation
- PB-GRPO: Learning Socially Adaptive LLM Agents from Persona-Driven Simulation with Preference-Batched GRPO
- ComPO: Community Preferences for Language Model Personalization