Do LLM Agents Respond to Where Social Proof Comes From?

People are more likely to adopt a message when the friends who already shared it come from separate social circles, not one tight clique. Network scientists call this structural diversity. On Sina Weibo, the probability of forwarding peaks at two exposures, and the most viral messages are dominated by exposures from two sources who do not know each other (Cumulative Effect in Information Diffusion). Simulations that use large language models as social agents now try to reproduce such spreading processes, the way an idea or behavior passes from person to person. Yet recent work shows that an LLM agent's willingness to adopt depends heavily on how its neighborhood is written into the prompt. Saying "3 of 10" rather than "30%" can flip whether the simulated spread looks like a disease or like a behavior that needs repeated reinforcement (One Crowd, Three Cascades). Those tests held only the count of active neighbors fixed. Nobody has asked whether an LLM agent responds to the arrangement of those neighbors, which is the variable the empirical literature says matters most.

We propose a controlled experiment on single agents. We hold fixed the number of active neighbors and vary three things independently: how those neighbors are connected to each other, how different their messages are, and how all of this is written into the prompt. An LLM agent "has" structural diversity only if two conditions hold. First, its adoption gap between diverse and clustered exposure must persist when diversity is conveyed implicitly and under equivalent phrasings. Second, that gap must not reduce to the agent reading more varied arguments. Agents that pass become a measured response function that can be checked against Weibo and plugged into classic questions about how spread crosses between communities. Agents that fail show where current LLM diffusion simulators produce artifacts.

Core Claim

The obvious combination is to tell an LLM agent "your three friends who shared this are from three different circles" and watch adoption rise. That experiment is easy and nearly uninformative, because it measures whether the model reacts to the phrase "different circles." Our claim is narrower and testable. A structural-diversity effect in an LLM agent is a social mechanism only if it meets three conditions. It must survive across semantically equivalent ways of encoding the active neighbors' relationships, which extends the invariance diagnostic of One Crowd, Three Cascades from counts to topology. It must appear when the agent has to infer diversity from relational descriptions or an edge list rather than from an explicit label. And it must remain once we control for the content diversity that disconnected sources naturally bring.

We write the agent's adoption probability as

$$\text{logit}\,P(\text{adopt}) = f(k) + \beta_s S + \beta_c C + \beta_{sc} S C + \gamma_e + \delta_e S,$$

where $k$ is the number of active neighbors, $S$ is the number of connected components among those active neighbors (structural diversity), $C$ is the measured semantic spread of their messages (content diversity), $\gamma_e$ is a fixed effect for serialization format $e$, and $\delta_e$ is how much the structural effect itself changes with format. The proposal's object of study is $\beta_s$ with $\delta_e \approx 0$. If instead $\delta_e$ dominates $\beta_s$, then the "diversity effect" in LLM simulations is typographical.

Where This Sits

Two seed papers frame the gap from opposite sides.

LLM-Augmented Agent-Based Simulation of Cyber Social Agents builds LLM agents from empirical bot archetypes and studies how persona diversity shapes diffusion. It does not separate that from the topological diversity of an agent's exposures, and it does not test thresholds. One Crowd, Three Cascades measures the exposure-response curve $P(\text{adopt}\mid k,d)$ and shows its serialization fragility. It serializes only flat aggregates and name lists, though, so topology is absent. We take that paper's method, the invariance diagnostic, and apply it to the variable it omits.

LAID lets GPT-3.5 agents rewrite messages as they pass them on. This supplies the content-diversity mechanism we must control for. However, LAID spreads messages edge by edge, like a simple contagion, and never compares clustered with disconnected exposures.

Lost in Serialization constrains the design in two ways. It shows that LLM graph reasoners break under edge shuffling and syntax changes, so the edge-list encoding of structure may fail outright. It also finds larger models more invariant, which motivates our model-scale axis.

On the network-science side, three papers motivate the target. An Empirical Evaluation Of Social Influence Metrics shows that the number of active communities predicts retweets better than raw neighbor count. Cumulative Effect finds the diverse two-source motif in highly viral messages. Toward Order-of-Magnitude Cascade Prediction ties tenfold cascade growth to community diversity among recently exposed users.

An Experimental Study of Structural Diversity is a constraint we take seriously. A 500,000-user randomized trial on Twitter found no retention effect from diversified follow recommendations. That measured long-run retention, not adoption of a given message. It still warns that observational diversity effects may reflect who people are rather than how they are connected. An LLM experiment can randomize topology for the agent, but it cannot settle the human question. We use human data only as a reference curve, not as a causal claim.

Research Program

Single-agent factorial experiments. For each scenario, a persona-grounded agent receives a message along with $k \in \{1,\dots,6\}$ active neighbors. The scenarios reuse the five-scenario style of the seed paper and add Grabowicz-style topical and non-topical content (Distinguishing Topical and Non-Topical Diffusion).

Structural diversity $S$ ranges from 1, where all active neighbors form one clique, to $k$, where they are mutually disconnected.

Content diversity has two levels. In the matched level, every neighbor's message is identical. In the varied level, each neighbor's message is a paraphrase or distinct framing produced by LAID-style alteration. The spread $C$ is measured by embedding distance. Crossing these two factors gives the disentangling cell no prior work has: diverse topology with identical text, and clustered topology with varied text.

Structure is serialized five ways:

  • an explicit community label per neighbor;
  • relational prose ("Ana and Ben have never met");
  • an edge list among active neighbors, in sorted and shuffled order;
  • relationship roles only (coworker, sibling, physician);
  • an implicit condition in which neighbors are named and the agent can retrieve their mutual ties through a memory lookup, but is never told.

All of these are crossed with the count, fraction, and enumeration encodings from the seed paper. We run small local models, matching the seed paper, and larger open models up to the scale tested in Lost in Serialization. Each model is run with and without explicit reasoning.

Reference curves from human data. From the Sina Weibo WISE 2012 data used in Cumulative Effect, we extract empirical forwarding rates by number of exposures and by whether the exposing sources are connected. The file reports the qualitative shape: a peak at two exposures and dominance of the diverse motif among viral messages. It does not report a numeric diverse-versus-connected adoption ratio. We will compute that ratio ourselves and treat it as the human target for the sign and rough size of $\beta_s$.

From single decisions to cascades. For every model and serialization that passes the invariance test, we fit the response function above. We use it as a cheap surrogate agent on the Facebook and arXiv graphs from Don't Be Greedy. On those graphs we ask two questions. Does the "wide bridge" bottleneck of complex contagion (Complex Contagions in Kleinberg's Small World Model) appear? Does community-aware seeding (dynamic-programming influence maximization, which spreads the seed budget across nested communities) still beat greedy seeding? We validate the surrogate against full LLM-agent cascades on community-preserving subgraphs of a few hundred nodes, which is the scale LAID-style simulation can afford.

Baselines

Count-only reference. The do-nothing reference is the count-only exposure-response curve of One Crowd, Three Cascades. Here the agent sees $k$ and $d$ but no structure, so by construction $S$ can have no effect. We run it with the same models, scenarios, personas, and sampling budget as every structural arm. Any structural arm must change adoption beyond this curve's confidence band at fixed $k$ to count as a structural effect.

Classical threshold models. The strongest classical competitors are fitted with the same maximum-likelihood procedure on the same Weibo exposure events. One is the Linear Threshold model. The other is a community-count threshold, where an agent adopts once enough distinct communities are active, mirroring the active-community-count feature of An Empirical Evaluation Of Social Influence Metrics. For predicting individual retweets, that paper's random forest reports precision 0.95 and F1 0.948, against F1 0.622 for its baseline, on the Weibo retweet benchmark. We do not expect a persona-prompted LLM agent to beat this, and we will not claim to. The comparison that matters is whether the LLM agent's fitted $\beta_s$ matches the sign and size implied by the community-count model on Weibo. An LLM agent "beats" the classical models only in the counterfactual regime they cannot represent: when content diversity and structural diversity come apart.

Cascade-level seeding. The strongest method is DPIM from Don't Be Greedy. Under the S-Cascade threshold model it reports a 55% average gain over greedy seeding and a 266% gain on Facebook; under independent cascade the lead is 7–8%. We run DPIM and greedy seeding with identical seed budgets under three response functions: the S-Cascade function, the count-only LLM surrogate, and the invariant structural LLM surrogate. The question is which regime the LLM surrogate's DPIM advantage falls into.

Key ablation. The ablation that isolates our ingredient is the content-matched, topology-varied arm against the content-varied, topology-matched arm. If only the second moves adoption, the agent responds to arguments, not structure.

Why not just…?

Why not just fit a classical threshold model with a community-count term and skip LLMs? For predicting real adoption from logs, this is the better tool, and we use it as the human reference. It cannot answer our question, though. Our question is whether LLM simulators, which are already being used to study diffusion of evolving text, carry the structural mechanism at all. A content-blind threshold model also cannot represent the case where disconnected sources send identical text (Bridging the Topological Semantic Gap), and that case is the decisive one.

Why not just state the community membership in the prompt? That is one of our five serializations. On its own it confounds structural reasoning with responding to the word "different." If the effect appears only with explicit labels and vanishes under relational prose, edge lists, and role cues, we will report that the agent follows a cue rather than represents topology. That is a useful negative result for anyone building diversity-aware simulators.

Why not just use a bigger model or chain-of-thought to remove serialization artifacts? This is a real threat to the framing. Lost in Serialization finds larger models more invariant, and the seed paper used only small models. We therefore treat scale and reasoning as experimental factors rather than assume the artifact persists. If large reasoning models show $\delta_e \approx 0$ and a stable $\beta_s$, the contribution becomes a positive certification protocol rather than a critique, and we will say so.

Why not just say diverse sources matter because they bring different arguments? This is the most plausible mechanistic rival, and it may well be right for LLMs. The factorial crossing of $S$ and $C$ tests it directly. If $\beta_s \approx 0$ while $\beta_c$ is large, structural-diversity effects in LLM simulations are effects of content diversity. That would be a substantive finding about what such simulators can and cannot represent.

Why not use a content-aware graph neural network? Such models are permutation-invariant and fuse text with topology. They are trained to predict from historical logs, however, and so cannot be queried in the randomized counterfactual cells, such as identical text from disconnected sources, that this study requires.

Proposed Experiments and Decisive Checks

The primary metric is the structural effect at fixed $k$: the adoption gap between $S=k$ and $S=1$. It is reported with bootstrap intervals for each model, scenario, and serialization. The secondary metric is an invariance ratio: the spread of that effect across serializations divided by its median magnitude.

The first decisive check is the invariance test. A model passes if its structural effect keeps its sign under all five structure encodings, including shuffled edge lists, and its invariance ratio stays below one.

The second is the content-separation test. A passing model shows $\beta_s$ reliably above zero in the identical-text arm.

The third is the human-alignment test. The sign of $\beta_s$, and whether adoption peaks at low $k$, should match the Weibo reference curve, including the two-exposure peak reported in Cumulative Effect. The magnitude target comes from our own extraction, because no numeric ratio is reported on Lacuna.

At the cascade level, the surrogate built from a passing agent should reproduce three things:

  • full-LLM final cascade sizes on subgraphs within their run-to-run variance;
  • the wide-bridge bottleneck on synthetic small-world graphs;
  • a DPIM-over-greedy advantage that sits between the independent-cascade and S-Cascade figures.

Agents that fail the invariance test should yield DPIM advantages that swing with serialization. That would extend the seed paper's warning from cascade size to seeding strategy.

The outcome is informative either way. Either we get the first measured, format-robust structural-diversity response for LLM agents, or we get a demonstration, model by model, that "diverse exposure" in these simulators is a prompt cue or a content effect rather than a social mechanism.

Workshop authors to connect with