WorksPhysics and Astronomy
Predictability of Collective Behavior in Networks of LLM Agents
Abstract
As AI agents built on large language models (LLMs) increasingly interact at scale, anticipating their collective behavior becomes necessary. Unlike humans, individual LLM agents can be probed in {controlled scenarios}; however, the predictability of their collective outcomes hinges on what such measurements capture. We study information cascades on networks of LLM agents that decide whether to share a post given how many of their contacts have shared it. At the same fraction of sharing contacts, an agent's response depends on how many contacts it has and on the post. Responses that ignore the number of contacts mispredict the cascades, increasingly on heterogeneous networks, by an error we derive analytically and that network topology alone cannot reveal. Measured in context, the response predicts, through mean-field theory, the cascades of full LLM simulations, tipping points included. Anticipating the collective behavior of AI agents thus requires evaluating them in their local network context.
Introduction
Large numbers of AI agents have started to act autonomously on our behalf, including shopping, searching for information, and posting on social media [1, 2]. LLM-driven AI agents interact not only with humans but with one another, a phenomenon one could recently witness in a number of cases ranging from social media platforms populated by AI agents [3, 4] to cybersecurity incidents caused by swarms of agentic AI systems [5]. It is natural to ask whether such populations reproduce the collective phenomena that human populations do [6], from the emergence of shared conventions and consensus [7, 8, 9] to polarization [10, 11] and the viral spreading of information [12, 13, 14, 15], and, crucially, under what circumstances and how precisely their collective behavior can be predicted. The question is of practical importance: when agents share content on our behalf, whether a post spreads is decided by the collective dynamics of the population, not by the choice of any single agent, and such an outcome should be anticipated before it happens rather than observed after the fact [16, 17, 18].
Predicting the behavior of a collective from that of its members is the classic micro-to-macro problem of the social sciences and of statistical physics [19, 20, 7, 21]. In human populations it is often hard to bridge this micro-macro gap, because the response of an individual to its social environment can rarely be measured directly: outside small controlled experiments [8, 22, 23], it must be inferred from observational data [24]. This obstacle does not exist for LLM agents. The response of a single agent to its (social) environment can be measured directly, by describing that environment to the agent in a prompt and querying the LLM as often as needed to estimate the probability distribution of answers. Each ingredient of the environment, from the number of contacts to the content the agent is shown, can then be varied on its own while all the others are held fixed, as in a controlled experiment, but with a precision and at a scale that experiments with humans cannot reach. Moreover, an LLM agent given the same prompt behaves the same in a measurement as in a population, whereas human behavior in the laboratory can differ from that outside it [25]. A promising wave of recent work has used this methodology to demonstrate that the collective behavior of LLM populations can indeed be predicted from the individual agent’s behavior, building on the observation that LLM populations spontaneously develop shared conventions [9]. The idea is to probe how a single agent responds to different environmental cues, and use mean-field theory to recover the macroscopic outcome, as demonstrated for majority-following and spontaneous coordination [26], for group-size effects in naming games [27], and for opinion dynamics with a fitted Ising-like response [28].
Most of these works, however, share a structural simplification. Agents are embedded in a well-mixed population, in which every agent can interact with every other, so that each agent responds, directly or through randomly drawn partners, to the state of the population as a whole, such as the average opinion [26] or the overall fraction of agents that have adopted a norm [27]. Recent work places the responses measured in this setting on networks [29], assuming that they do not depend on the agent’s position in the network. In realistic settings, by contrast, LLM agents will often have access to information about their immediate environment, such as what the accounts they follow have posted or shared, and this local information may shape their behavior more than any global signal. If it does, the network in which the agents are embedded becomes decisive for the collective dynamics, because it determines the information each agent is exposed to. In many collective phenomena, from the adoption of a convention to the sharing of a post, an agent’s decision may then depend on how many of its own contacts have already adopted a behavior. Since agents on a network differ in how many contacts they have, it matters whether an agent responds to the fraction or to the number of these contacts. Threshold models of contagion have long settled this question by assumption [20, 30, 31, 32]; for LLM agents it can be measured directly. Experiments in well-mixed groups already show that how strongly LLM agents follow a given majority decreases as the group grows [26].
This raises a general question about populations of LLM agents. How is the behavior of a single agent linked to that of the population, and which aspects of a single agent’s response must be measured to predict the collective behavior? In principle, the response of an LLM agent can depend on everything in its prompt, from how many of its contacts have shared and how many contacts it has to the content of the post, its memory of past interactions [33, 31, 9], the persona or ideology it is given [34, 11] and the incentives of the agent, such as reward maximization [35], as well as on the language model [36] and its biases [37, 38]. Each variable left out of the measurement is a potential source of error in the prediction. We address the question for information cascades, for which the theory linking individual responses to collective outcomes on networks is mature [30, 31, 39, 40], so that comparing the prediction with full LLM simulations tests the measured response rather than the theory. We measure the response as a function of the most structural of these variables, how many contacts an agent has and how many of them have shared, for different posts, and keep the others fixed.
We show that the collective behavior of interacting LLM agents can be predicted from measurements of a single agent, provided the measurement captures the agent’s local social context. We study information cascades on social networks of LLM agents, in which each agent decides whether to share a post after seeing how many of the accounts it follows have shared it. First, we find that networked LLM populations produce information cascades whose reach depends on both the network and the post. The decision of each agent depends on its local context in a way that no single response curve can describe. Agents who follow more accounts share when a smaller fraction of them has shared: an agent that sees 2 of its 10 contacts share a post almost never shares it, while one that sees 20 of its 100 contacts share it does so far more often. The response also changes with the content of the post. Second, we show that responses that ignore the number of contacts mispredict cascades, increasingly so on more heterogeneous networks and in a direction that depends on the content, whether they depend on the fraction of sharing contacts alone, as threshold models assume, or are measured in a single large group, as in previous studies. Third, the response measured in context, fed into the mean-field theory of cascades on networks, predicts the cascades of full LLM simulations, including whether they spread through the network and the seed fraction at which they tip; once the response is measured, the cascade on any network follows without further queries to the LLM. This agreement is expected for agents that decide only on what they are shown, provided the measurement records what they respond to; the failure of the responses that ignore the number of contacts shows that this condition is not met by default. Finally, we derive the error of these responses analytically, for any response that depends on the number of contacts, and show that it cannot be inferred from network statistics alone. The collective behavior of LLM populations is thus predictable, but only if each agent is measured in the social context in which it acts.
Results
Information cascades in networked LLM populations
Exposure-triggered sharing dynamics We simulate the spreading of four different social media posts on a social network of LLM agents; the process is schematically depicted in Figure 1. Specifically, we study four posts on contested political topics: two posts about migration, one about defunding the police and one about gun rights (Figure 2). We chose them because the spread of such posts on social media matters [41].
Figure 1. Information cascades in a population of LLM agents. (A to D) Spreading from the seed agents () to the end of the cascade (), when no agent shares any more. (E) A single decision: the prompt tells the agent how many of the accounts it follows have shared the post, and the agent shares it (YES) or not (NO); an agent that declines is asked again only if more of its contacts share.

Figure 2. Information cascades in LLM populations depend on both the network and the post. Final cascade size against the seed fraction in the LLM population, an LLM agent at every node, on Erdős–Rényi (ER) networks with mean degree and and on Barabási–Albert (BA) networks, all with agents. (A to D) The four posts; each panel gives the text of the post and the name used in the text. For three posts the cascade spreads most easily on BA and least on the sparse ER network; for gun-right the sparse ER network is the one on which the smallest seed already triggers a large cascade. Each point averages 100 runs; the error bars (standard errors) are mostly smaller than the symbols.
The LLM that governs the decisions of our agents is the open-weight model mistral:7b; it gives us full control over the prompt and the decoding, and makes every measurement reproducible.
We consider a network of LLM agents, one at each node of the social network, where each agent follows its neighbors. At the start of the dynamics, a fraction of agents, chosen uniformly at random, share the post (seed agents). The cascade then unfolds in discrete rounds. An agent decides whether to share the post whenever something changes in its local environment, i.e., when the number of its neighbors sharing the post has increased since it last made a sharing decision. Agents with no sharing neighbors never encounter the post, and therefore never get to make a sharing decision. An agent that declines is not excluded from the process: It may reconsider once further neighbors share the post, so repeated exposure gives a post several opportunities to be adopted. This mirrors how content reaches users on social media, where a post can reach a user again each time another account they follow shares it. All decisions within a round are made simultaneously, based on the state of the network at the beginning of that round. Newly sharing agents then raise the exposure of their own neighbors, who are asked in the following round (see Figure 1). The cascade ends when a round produces no new sharers. We record the final fraction of sharers, , as the size of the cascade.
Each decision of an LLM agent is a single, independent query to the language model (Materials and Methods). The model is instructed to act as a user on a social media platform that tries to maximize engagement. It is shown the post and told that the post is supported by out of the peers it follows, and it is asked to answer YES or NO to whether it shares; the prompt is depicted in Figure 1E. Queries carry no memory: an agent receives no information about its own earlier decisions or about the history of the cascade beyond the current count . Replies are parsed for YES or NO, and a reply that contains neither is handled as described in Materials and Methods.
Interplay of network topology and post content The results of the full LLM simulations are shown in Figure 2. Each network has agents, and for each network and post we run 100 cascades at each of 12 seed fractions (Materials and Methods). Both the final cascade size and the seed fraction required to trigger a global cascade depend strongly on the network structure and on the content of the post. We call a cascade global when it reaches at least half of the agents, and the tipping seed the seed fraction at which the cascade becomes global. The effect of the network can be judged within each panel, where the different symbols (and colors) correspond to different network structures. Generally, global cascades are triggered more easily in Barabási–Albert networks than in Erdős–Rényi networks of comparable mean degree, and within Erdős–Rényi networks they become easier to trigger as the mean degree increases, except for gun-right, whose cascade reaches half of the agents from a smaller seed on ER () than on ER (0.06). The content of the post, depicted in panels A to D, also introduces considerable variation in the seed fraction needed to trigger a global cascade: across posts it varies by a factor of 2.6 on BA (from 0.056 to 0.147) and 3.5 on ER (from 0.059 to 0.206).
The response of an agent depends on its degree and the post content
Measuring the response of a single agent To better understand, and ultimately predict, information cascades in realistic LLM populations, we construct reduced models of the system. To this end, we first measure the sharing statistics of individual LLM agents in terms of a function , defined as
the accounts an agent follows are its neighbors on the network. The idea is to replace LLM agents (making sharing decisions about posts) with simpler probabilistic agents that have the same information-sharing statistics. Since each query contains only the post, and , measured for each post determines the decisions of the agent in our setting. The same measurement can be conditioned on any further variable an agent may respond to, such as its memory of past interactions, the persona or ideology it is given, or the model it runs; here we focus on the variables and . The function is a probabilistic (“soft”) response: it can take any value between 0 and 1, rather than jumping from 0 to 1 at a threshold. We measure by presenting individual agents with exactly the prompts used in the full population simulations (Figure 1E). For each combination of post, and , we query the LLM 200 times and estimate as the fraction of YES responses, i.e., the probability that an agent shares the post in a single decision when of its followed peers have shared it. The response is measured for 147 degrees between and , which includes every degree that occurs in our networks (the largest is 155, on BA), so it is never used beyond the degrees at which it was measured.
The response depends on degree Figure 3A shows the measured response functions for each post and for several values of . Our first and most important observation is that does not depend on the ratio alone: the curves for different do not collapse onto a single curve.3 Instead, depends strongly on the degree , and for higher-degree agents it typically saturates at a smaller fraction of sharing peers. Second, increases monotonically with only for small degrees; for larger it becomes non-monotonic. Third, the response functions differ considerably between posts. For example, at most degrees, more social pressure (a larger ) is needed to raise the sharing probability of “Migration must be stopped.” than that of “Migration must be limited.”

Figure 3. The measured response predicts the cascades; its degree average does not. Rows: the four posts, as in Figure 2. (A) Measured probability of sharing against the fraction of sharing contacts, one curve per degree , for 14 degrees between and spaced logarithmically (color scale at the top); orange: the degree-averaged response of the ER network. (B to D) Final cascade size against the seed fraction on ER (continuous transition), ER (discontinuous) and BA, all with . Lines: mean-field theory with the degree-resolved response (black) and with the degree-averaged response (orange, dashed). Symbols: Monte Carlo averages with the same two responses (black circles, orange triangles), single Monte Carlo runs (gray dots) and the LLM population (blue squares; error bars: standard errors). The theory matches the Monte Carlo simulations for both responses on every network, and the LLM population follows the Monte Carlo simulations with , except for a slightly earlier spread of police-defunded and, to a lesser extent, of migration-must-be-limited. With , the first three posts ignite at a lower seed on both ER networks, and from a very small seed on BA. For gun-right on ER the direction reverses: only the cascade with tips (§ 2.2).
For all four posts, the probability of sharing depends strongly on the social signal: it is close to zero when only a small fraction of the agent’s contacts has shared, and close to one when all have (Figure 3A), so the agents do take into account what their contacts did, and the probability usually rises gradually rather than in a single step. The non-monotonicity at large is a small modulation of this overall increase with ; in any case, the analysis presented below does not require any particular shape of this response: the theory of § 2.3 applies to any measured , monotone or not.
Low-degree agents (dark) need a higher fraction to reach , high-degree agents (bright) a lower one, so the threshold fraction drifts downward with degree. (We define as the fraction at which reaches : the midpoint of the response, where sharing becomes more likely than not.) At
large degree this drift is common to all four posts (– at , – at and – at ); at the smallest degrees it is not, as we discuss in § 2.2. The response is therefore neither a pure fraction rule ( constant) nor a pure count rule. Instead, we find a fraction-like behavior with a systematic degree drift, where higher-degree agents tend to ignite at a lower fraction. This agrees with De Marzo et al., who found that how strongly LLM agents follow a majority depends on the size of the group [26].
The response depends on the content The drift with degree is not the only systematic feature of the response. With the same model, the same prompt template and the same degree, the response also changes with the content of the post, and it changes most where the social signal is weakest. Consider an agent that follows three accounts, one of which has shared the post. It shares gun-right (“Everybody should have the right to own a gun.”) with probability , but each of the other three posts with probability at most . At large degree, by contrast, the four posts are alike (Table S1, § 6.1). This difference at low degree decides the collective outcome on a sparse network, where most agents follow few accounts. There, a cascade can grow from a small seed only if a single sharing contact is often enough to make an agent share, which is Watts’ cascade condition [30]. On ER this condition nearly holds for gun-right and is far from holding for the other three posts (§ 6.1). Accordingly, gun-right is the only post whose cascade on this network jumps, from a seed of to a final size of , while the cascades of the other three grow smoothly with the seed (Figure 3B).
| post | |||
| “Migration must be limited.” | 0.41 | 0.20 | 0.29 |
| “Migration must be stopped.” | 0.04 | 0.02 | 0.34 |
| “The police should be defunded.” | 0.69 | 0.23 | 0.32 |
| “Everybody should have the right to own a gun.” | 1.00 | 0.82 | 0.28 |
Table S1. Where the four posts differ. : probability of sharing when one of the accounts followed has shared (measurement uncertainty ). : threshold fraction at , a large degree, inside the measured range, at which the four posts agree.
Predicting the cascades requires the response measured with the degree
Previous studies give every agent the same response Measuring the degree-resolved response requires querying the model for every number of sharing contacts at every degree that occurs in the network. Previous studies that predict the collective behavior of LLM agents from the response of single agents [26, 27] measure a single response function in populations where all agents have the same number of contacts (a fully connected group, in which every agent sees all the others, or pairs of agents drawn at random) and apply it to every agent. A response that depends on the fraction alone gives an agent with 2 sharing contacts out of 10 the same probability as one with 20 out of 100; the measured response gives at most 0.01 to the first and between 0.20 and 0.94 to the second, depending on the post (Figure 3A). On networks, where agents differ in how many contacts they have, we show that the prediction of the cascade fails without quantifying the degree dependence. This setting tests which aspects of a single agent’s response must be measured to predict the collective behavior.
Predicting a cascade from a single-agent response A measured response predicts a cascade in two ways. In Monte Carlo simulations, each LLM query is replaced by a random draw from the response, on the same networks and with the same update rule as in the LLM populations (Materials and Methods). With the empirically measured at hand, we can go one step further: instead of performing Monte Carlo simulations of the cascades, we can predict them analytically using the tree-based mean-field theory of Gleeson and Cahalane [39], which assumes that the network is locally tree-like and that the degrees of neighboring agents are uncorrelated, conditions that our networks nearly satisfy (§ 6.2). This gives both an efficient way to compute the final cascade size on any network and analytical insight into why cascades in LLM populations ignite where they do.
The theory takes as input a response function : the probability that an agent of degree with active neighbors is itself active. The crucial step is therefore to relate the measured single-decision probability to this response. The two are not the same, because an agent is asked again each time its exposure () grows. An agent that ends the cascade with sharing neighbors has been asked once at each exposure level , or until it shared. Since its decisions are independent, the probability that it has shared is one minus the probability that it declined every time:
Two properties make the appropriate input to the theory. First, it depends only on the final count and the degree , not on the order in which the neighbors shared. Second, it never decreases with , since every further exposure is one more chance to share, regardless of the shape of (§ 6.2). Equation (2) thus links the microscopic decision we measure from the LLM to the macroscopic cascade on the network. Equivalently, each agent behaves as if it drew a hidden threshold once, from the distribution , and shared as soon as reached it (§ 6.2). Since depends on and separately, and not only on the fraction , our model is an instance of the generalized, degree-dependent threshold model [31, 42, 43], rather than the fractional-threshold model [30].
The mean-field theory holds for any response that is non-decreasing in ; we write it for general , because the same equations will be used to quantify the error of the degree-averaged response (§ 2.5). An agent of degree whose neighbors are each active with probability becomes active with probability , the average of over the binomial number of active neighbors among the neighbors other than the one through which the agent is reached (§ 6.3). The probability that a randomly chosen edge leads to an active agent obeys the self-consistency equation of Gleeson and Cahalane [39], and its solution gives the final cascade size:
with the degree distribution, the excess-degree distribution, and the probability that an agent of degree is active, given that each of its neighbors is active with probability (the same binomial average, over all neighbors). We take to be the smallest fixed point of , obtained by iterating from : the cascade grows from the seeds and stops at the first fixed point it reaches, even if larger ones exist. Note that , , and are all linear in the response , a property we exploit below.
Averaging over degrees mispredicts the cascade We compare the cascade predicted with the measured response with those predicted by two degree-blind responses: (i) the average of our measured responses over the degrees of the network and (ii) the response measured in a single large group.
The first approximation is built from the same measured responses, discarding only the information about degree, so that the comparison isolates the effect of ignoring the degree. It is the degree-averaged response , the plain average, over all the degrees that occur in the network, of the measured responses read as functions of the fraction of sharing contacts (Materials and Methods), and it gives an agent with sharing contacts out of the probability , whatever its degree. Through Eq. (2), the cascade with still depends on degree, because an agent with more contacts has more chances to share before a given fraction of them has shared; what removes is the dependence of each decision on degree. In Figure 3A, is the thick orange curve (shown for ER ); it disregards the spread of the colored curves around it. We use the plain average because it is the simplest choice. Weighting each degree by how many agents have it or by how often a cascade reaches it, as a measurement on randomly chosen agents or on the agents that a cascade reaches would do, does not remove the error: it is still at least 0.029 in every combination of post and network (§ 6.4).
In Figure 3, orange marks the degree average: itself in panel A (for ER ), and in panels B to D the cascades with , from the mean-field theory (dashed lines) and the Monte Carlo simulations (triangles). The theory matches the Monte Carlo simulations with , as it matches those with , for every post and network, so the gap between the two cascades is the error of the degree-averaged response alone. We write this error as , the final size of the cascade with minus that of the cascade with at the same seed fraction : it is negative where averaging overpredicts the cascade and positive where it underpredicts it. For three of the four posts — the two migration posts and police-defunded, in the first three rows of Figure 3 — the degree-blind ignites every network at a lower seed (earlier tipping on ER and BA, uniformly higher on ER ). The fourth post, gun-right (bottom row), behaves in the same way on ER and BA but not on ER (see below).
The overprediction, for these three posts on every network and for gun-right on ER and BA, follows from how the response depends on the degree. The threshold fraction , the fraction of sharing contacts at which the response of an agent of degree reaches (§ 2.2), falls with degree (Figure 3A). The degree average gives the same weight to every degree that occurs in the network, so its threshold is lower than that of the typical agent, which has few contacts, and the cascade with spreads more than the cascade with . On BA, where most agents have 3 to 5 contacts but 147 different degrees occur, the degrees , which only 11% of the agents have, carry 94% of the weight in , and the cascade with ignites the network from a vanishingly small seed (§ 6.5). The networks sit in different regimes. On ER the transition is continuous, except for gun-right (§ 2.2): the final size grows smoothly with the seed fraction. On ER and BA it is discontinuous: the final size stays small below the tipping seed and jumps to a global cascade there (Figure 3B to D).
The exception is gun-right on ER . On this sparse network the cascade grows through agents with few contacts, whose response to a single sharer is high for this post (§ 2.2). The degree average underestimates this response: at the fraction it gives , while the measured is 0.82. Averaging therefore underpredicts the final size, by up to , and misses the global cascade of the LLM population. On ER and BA the cascade grows through agents of higher degree, where the four posts respond alike, and the degree average overpredicts the cascade of gun-right as it does for the other posts. § 2.5 shows that, to first order, the error has the sign of the response deviation where the cascade samples the response.
A large-group response mispredicts the cascade as well The second approximation mimics a measurement in a large group in which every agent sees all the others, as in De Marzo et al. [26]. In our measurement an agent with contacts sees others, so our largest degree, , is the largest such group we have; we give its response to every agent. This large-group response fails, in opposite directions depending on how it is read: read at the fraction , it ignites every network from the smallest seed; read at the count , it hardly spreads beyond the seed. For every post and network, with either reading, its error, the mean of over the seed fractions, is at least as large as that of , with one exception, gun-right on ER read at the fraction, where the agreement is accidental. Because it gives every agent the response of a degree that almost none of them has, it fails even when every agent has the same degree, where is exact (mean-field theory, § 6.6).
The response measured with the degree reproduces the LLM populations The Monte Carlo simulations with the measured (black points in Figure 3B–D) closely reproduce the cascades of the full LLM populations (blue squares). This agreement is expected: each agent’s decision is binary, independent and memoryless, and depends only on the post, and . It is therefore a Bernoulli trial fully characterized by , which captures not just the mean of the sharing behavior but its entire distribution. Cascades on networks of LLM agents can therefore be predicted from single-agent measurements alone, but only if the measurement records every variable the agent responds to: here the degree as well as the number of sharing contacts, which the degree-blind responses above leave out.
The LLM population follows the Monte Carlo simulations with on every network, except that police-defunded and, to a lesser extent, migration-must-be-limited spread slightly earlier than their responses predict. On the two ER networks, the LLM population of police-defunded lies closer to the cascade with than to that with .
Fed with , the theory reproduces the behavior of the full LLM populations on all networks and for all posts with good accuracy (black lines in Figure 3B–D). The deviations close to the tipping points, especially on scale-free networks, are finite-size effects that smooth out the transition.
The error of the degree average grows with network heterogeneity
The error of the degree average differs widely between posts and networks (Figure 3). This section shows what sets its size: how much the response depends on degree, and how heterogeneous the network is.
The error grows with the degree-dependence of the response The error of the degree average should grow with how far the degree-resolved response lies from it. At each count and degree , this distance is , the vertical gap between a colored curve and the orange curve in Figure 3A. The cascade, however, depends on the response through the effective response of eq. (2), so we take the gap there, ; to first order, the cascade error is a weighted sum of (§ 2.5). To compare posts and networks, we condense into one number, the degree dependence of the response . It is the root-mean-square of over and , with each degree weighted by , the probability that a neighbor of an agent has degree , because a cascade reaches agents through their neighbors (§ 4.3; § 6.7 explains this choice). Figure 4A plots the averaging error against this number for every post and network, with one line per post joining the three networks.
Figure 4. The error of the degree average grows with the degree dependence of the response and with the heterogeneity of the network. (A) Averaging error against the degree dependence of the response , from the Monte Carlo simulations (25 seed fractions). Color: post; marker: network; the lines join the three networks of each post. The error grows along the line of every post and is largest on BA. (B) Averaging error against the degree heterogeneity at fixed mean degree , from the mean-field theory (601 seed fractions). The networks are obtained by rewiring a regular graph, on which (top axis: rewiring steps per agent, ). The error is zero on the regular graph and grows with heterogeneity for every post.
The error grows with the degree dependence of the response along the line of every post, and every post has its largest error on BA, where that dependence is largest (–). The points do not, however, fall on one curve, as they would if alone determined the error. On ER , for instance, three posts have almost the same degree dependence ( between 0.087 and 0.090), but their errors differ by a factor of three: 0.037 for gun-right, 0.073 for migration-must-be-stopped and 0.113 for migration-must-be-limited. The error depends not only on how much the response depends on degree, but also on where, at which counts and degrees, because the cascade samples the response only at some of them (§ 2.5). It therefore cannot be forecast from , or from any other single number that summarizes the response.
Heterogeneity at fixed mean degree We now test how the error depends on network heterogeneity. The networks of Figure 4A differ in both mean degree and heterogeneity. To separate the two, we start from a regular graph with and rewire it step by step [44], which makes the degrees more unequal at fixed mean degree ( denotes the number of rewiring steps per agent), and we compute the error with the mean-field theory (Materials and Methods), which reproduces the Monte Carlo simulations for both responses on every network (Figure 3). Heterogeneity is measured by . On the regular graph () the error is zero (Figure 4B), because a network with a single degree leaves nothing to average over. As the rewiring makes the network more heterogeneous, the error grows for all four posts, reaching 0.17–0.30 at the largest heterogeneity (, ). The growth is monotonic for every post, although slow for gun-right between and 1. § 2.5 explains this growth analytically.
At , 9 and 12 the error also grows with heterogeneity (Figure S1): the rank correlation between the error and is positive in all sixteen sweeps (four posts at four mean degrees), between 0.89 and 1.00. We draw no conclusion about how the error depends on the mean degree itself, because that dependence comes from how is built: the plain average gives every degree that occurs the same weight, and a denser network spans more degrees, so its average contains more of the permissive high-degree responses. With the averages that weight each degree by how many agents have it or by how often a cascade reaches it, the trend reverses (Section S8).
Figure S1. The averaging error grows with heterogeneity at every mean degree. Averaging error against the degree heterogeneity , from the mean-field theory (601 seed fractions), on networks rewired from a regular graph with mean degree (A), 7 (B, as in Figure 4B), 9 (C) and 12 (D). Points: rewiring steps per agent. The error is zero on the regular graph, and its rank correlation with is positive in all sixteen sweeps, between 0.89 and 1.00.
Heterogeneity therefore controls the cost of degree-averaging: the error vanishes on a homogeneous network and grows as the degrees become more unequal, for every post and every mean degree we tested.
Predicting the averaging error requires measuring the degree dependence
The degree dependence cannot be corrected for If the error of the degree-averaged response could be predicted from the network and from a single number summarizing how much the response depends on degree, a degree-blind response corrected by a known factor would suffice, and the degree dependence could be left unmeasured. We test this in two steps. First, we derive the error analytically: we expand the cascade theory of § 2.3 [39] around the degree-averaged response, which gives, to first order and in closed form, the error of any degree-blind response. Second, we test whether this error factorizes into network statistics times a single number for the response. It does not: the error depends on the full shape of the degree dependence, which must therefore be measured, not corrected for.
The error is a weighted sum of the response deviation Write the two responses that feed the theory as (degree-resolved) and (degree-averaged), and their difference . We expand the error around the solution with , the smallest fixed point of . Because the equations of the theory are linear in the response, to first order in the error in the final cascade size is a weighted sum of the response deviation (Section S9):
where is the binomial probability of active neighbors out of , and are the derivatives of and with respect to , and is the susceptibility. The susceptibility measures how strongly the cascade amplifies a small change in the response: a link that leads to an extra active agent makes, on average, further links do so, and sums this chain. The first term is the direct change in the final size; the second is the change in how activity propagates along edges, amplified by . We tested this first-order expansion against the Monte Carlo simulations (Section S10). It holds where the two cascades stay close: where both grow smoothly with the seed fraction, as on ER , and where both jump at nearby tipping seeds, as on ER , for which it gives the shift of the tipping seed. There it reproduces the measured error within 20%, except in the case with the largest error, migration-must-be-limited on ER , which it overestimates by about a third: as expected for a first-order expansion, the overestimate grows with the size of the error. It fails where, over a wide range of seeds, one cascade has spread to most of the network and the other has not: on BA, where the cascade with spreads from a vanishingly small seed and the cascade with only from 0.04–0.13, and for gun-right on ER , whose cascade with jumps at .
Three properties of the weight explain the errors of § 2.3 and § 2.4. First, is never negative, so the error has the sign of where is large, that is, where the cascade samples the response: when the degree-averaged response is too permissive there (), the average overpredicts, as for three of our four posts on every network and for gun-right on all but the sparsest one. Because this sign depends on the post, no single correction factor can account for it. Second, is a binomial peaked at , the counts the cascade samples, and its propagated part is tilted toward hubs by , so the error grows with degree heterogeneity. Third, diverges at the tipping point, where , so the deviation is amplified most there. The same expansion gives the shift of the tipping seed as another weighted sum of (Section S9); it is positive in all our discontinuous cases, and the degree-blind response underestimates how hard the network is to ignite.
The error does not factorize into network and response The error could be predicted without measuring the degree dependence if it factorized into a property of the network times a single number for the response, error . Here is an amplification set by the network alone, through its mean degree and the squared coefficient of variation of its degrees, , and is a single number for how strongly the response depends on degree. A typical value of could then replace the measurement, and degree averaging would be a controlled approximation.
The test needs a number that describes the response alone, because it asks whether the error separates into a part set by the network and a part set by the response. The degree dependence of Figure 4A does not qualify: it is computed from the degree average of each network and weighted by the of that network, so it changes from network to network. We therefore take to be the root-mean-square departure of the measured response from a pure fraction rule, which depends only on the post (Section S9).
We test the factorization on the heterogeneity sweeps of § 2.4 (Figure S1). If it held, the ratio error/ would be the same for all posts on a given network. Instead, it differs between posts by a factor of about two (Section S9), although hardly differs between them (0.13–0.15). The reason is the weight of eq. (4): it gives the deviation of the response the most weight where the cascade samples it, at counts near , at degrees weighted toward hubs, and near the tipping point. Two responses with the same can deviate in different places, and a single number cannot tell them apart. The network thus sets the scale of the error, but not its size for a given post. Degree averaging is therefore not a controlled approximation on heterogeneous networks, and the degree-resolved response must be measured rather than summarized.
Discussion
The collective behavior of networked LLM agents can be predicted from measurements of a single agent, provided the agent is measured in the local context in which it acts. In our networks, each agent decides whether to share a post after seeing how many of the accounts it follows have shared it. Populations of such agents produce information cascades whose reach depends on the network and on the post. How an agent responds depends on how many contacts it has: the more contacts it has, the lower the fraction of them that must have shared before sharing becomes more likely than not. The post matters mainly when the agent has few contacts. The response of a single agent to sharing contacts out of , measured at every degree and fed into the mean-field theory of cascades, predicts the cascades of the full LLM simulations, including whether they spread and the seed fraction at which they tip. Measured without this context, averaged over degree or taken from a single large group, the response mispredicts the cascades, increasingly so on heterogeneous networks and in a direction set by the content of the post.
These results bear on what a measurement of a single agent must capture. Predicting the collective behavior of LLM populations from the behavior of single agents has succeeded in several settings [26, 27, 28], most of them well-mixed groups in which all agents face the same situation. Where agents face different situations, the prediction holds only if the measurement resolves them, because the response is a property of the agent in its context, not of the agent alone. On a social network the context includes the number of contacts. Our agents follow neither of the two rules assumed by classic threshold models, a fixed fraction or a fixed number of sharing contacts [20, 30, 31, 32]: their threshold fraction falls with their degree, in line with the weakening of majority-following with group size in well-mixed groups [26]. In human populations the evidence is mixed: the adoption thresholds of Skype users, measured as fractions of their contacts, appear not to depend on degree [45], whereas on Twitter highly connected users are less likely to share any given message, because their attention is divided among more incoming messages [46, 47]. A response measured out of context mispredicts the collective outcome, in line with the observation that the collective biases of LLM populations are not easily deduced from agents in isolation [9]. The network thus shapes the collective outcome not only through who interacts with whom [29], but also through the local context it creates for each agent, which changes how the agent responds.
Measuring the local response properly, and therefore predicting the collective behavior, is not trivial. The response must be known at the degrees that occur in the network, but the structure of real social networks is rarely known in full, and it changes over time, as accounts join and leave and links form and dissolve [48, 49]. A measurement must therefore anticipate degrees not yet observed. Covering every degree is costly: an agent with contacts can see nonzero counts of sharing contacts, each needing its own queries, so the cost grows with the square of the largest degree; up to we made about two million queries per post. Unmeasured degrees must be interpolated or extrapolated, and whether the response of an LLM agent is regular enough for this is open. On heterogeneous networks the degrees hardest to cover, those of agents with many contacts, also carry more weight than their share of the population: in the error derived in § 2.5, the part that propagates along edges weights the response of each degree by , the probability that a neighbor has that degree. A measurement made for one network thus carries over to another only if it covers that network’s degrees.
We derived the error of the degree-averaged response analytically. To first order, it is a sum over the pairs of the deviation of the response from its degree average, each weighted by a factor set by the network and by the cascade with the degree-averaged response (eq. (4)). This form of the error has a practical consequence: a prediction made with a degree-blind response cannot be corrected by a factor computed from the statistics of the network. The error does not factorize into a property of the network times a single number for the response, because the weight selects where the deviation matters, at the pairs that the cascade samples, and posts whose responses are equally degree-dependent differ there. On the same networks, the error per unit of degree dependence varies between posts by a factor of about two § 2.5. The degree dependence of the response therefore has to be measured, not corrected for. The weight also shows where the measurement must be most precise. It is amplified near the tipping point, where the susceptibility diverges, so whether a post spreads is the prediction most sensitive to errors in the measured response.
Our conclusions rest on one language model, one prompt and four posts, and our agents have only a minimal persona and no memory. The response to gun-right differs from that to the other three posts at low degree, which sets its cascade on the sparse network apart (§ 2.2), but whether this difference survives rewordings of the post has not been tested. The LLM population follows its measured response closely, but for police-defunded and, to a lesser extent, for migration-must-be-limited it spreads slightly earlier than the response predicts (§ 2.3). The expansion of the error is first order: on ER it overestimates the largest error by up to about a third, for migration-must-be-limited (Section S10), and the expansion fails where the two cascades do not stay close, on BA and for gun-right on ER , where the error must be computed from the two mean-field solutions. Finally, our networks are synthetic and locally tree-like, and on clustered networks the tree theory loses accuracy.
Our agents see only how many of their contacts have shared, so the measured response describes them completely, and the full LLM simulations follow it closely. Agents deployed in practice carry more: a more complex persona or an ideology [34, 37, 11], a memory of past interactions [33, 9, 27], and the content of what their contacts shared and the identity of who shared it [38, 36]. Each of these makes the response depend on more than . A persona makes the response differ from agent to agent; memory makes the successive decisions of an agent depend on each other, so that its effective response no longer follows from by eq. (2); and the arguments attached to a message can sway an agent more than the identity of its sender [36]. For such agents the response can be refined step by step, from a single to one response for each persona and then to a policy that depends on the history of the agent, until the predicted and the simulated collective behavior agree. The gap that remains at each step measures how predictable a given design of agents is, and the step at which it closes identifies the minimal description of the agent that the collective behavior requires.
A measured response also indicates where to intervene. The theory extends directly to seed fractions that depend on degree, so it predicts how much the tipping seed falls when the first sharers are hubs, a placement with which far fewer committed agents overturn a convention in networked LLM populations [29]. The weight shows whose response matters most for the collective outcome, on heterogeneous networks that of the hubs, and therefore which agents an audit or a change of instructions should target first. The way a platform displays social proof, as a count, as a fraction or not at all, changes the response itself, and the same pipeline can measure how each choice changes the collective outcome.
The approach extends to clustered networks for which cascade theories exist [50], to networks that co-evolve with the behavior of the agents [49, 11], and to reversible dynamics such as consensus and polarization [7, 10], whose mean-field theories also take a response resolved by degree [40]. Because a measured response predicts the collective outcome on any network whose degrees it covers, without further queries to the model, responses measured for each model, once their predictions have been tested on real-world social networks, could anticipate the collective behavior of a population of agents before it is deployed [16], including populations that mix models, which need one response per model [36].
Materials and Methods
The LLM agents and the measurement of the response
Every query, in the measurement of the response and in the full LLM simulations, uses the following prompt, in which {m} and {k} are replaced by the number of the agent’s neighbors who have shared the post and by its degree:
You are a user on a social media platform that tries to maximize engagement.
A post by another user is shown below.
Post: "..."
The post is supported by {m} out of {k} of the peers you follow.
Do you support the post to reshare it? Answer with exactly one word: YES or NO.
Each element of the prompt serves the measurement. The persona, a user of a social media platform that tries to maximize engagement, gives the agent a reason to share; the count “ out of ” lets it respond to the number of sharers, to their fraction or to both, so that the response can depend on degree; and the one-word answer makes each query a binary decision. Because queries are stateless and show only this count, the response depends on alone, as the mean-field theory assumes. The prompt treats sharing as support: it describes the contacts who have shared the post as supporting it. We measure the response of § 2.2 for the LLM mistral:7b (via Ollama, ) and each post. For each pair we make an independent set of 200 calls, with the same prompt as in the full LLM simulations, and is the fraction of YES answers; its standard error, , is at most 0.035. If the model answers neither YES nor NO but declines to take a position, the query is repeated, and after five such answers in a row the decision is drawn at random, YES or NO with equal probability. The same rule applies in the full LLM simulations. We draw the decision at random, rather than counting a non-answer as NO, so as not to bias the response toward sharing or toward not sharing.
Full LLM and Monte Carlo simulations
We run the exposure-triggered cascade on three kinds of network, all with agents: Erdős–Rényi (ER) graphs [51] with mean degree and , and Barabási–Albert (BA) graphs [52], in which every new agent links to three existing ones. On ER, is sparse enough that the final size grows continuously with the seed, while is dense enough for a discontinuous jump, so the two cover both kinds of transition. BA has a mean degree close to 6, near that of ER , so going from ER to BA changes mainly the heterogeneity of the degrees. For each network we generate 500 graphs. Each Monte Carlo cascade starts from a fraction of randomly chosen seed agents, for 25 values of between 0.005 and 0.30; this window extends beyond the largest tipping seed of the degree-resolved response (0.21), so that every jump is captured. For each seed fraction we run 500 cascades, on the 500 graphs, and average their final size. The LLM population, with an LLM query at every exposure, runs on the first 100 of these graphs: 100 cascades at each of 12 seed fractions between 0.02 and 0.26 (9 seed fractions, up to 0.19, for gun-right on BA). The median standard error of the average final size is 0.0003 for the Monte Carlo simulations and 0.001 for the LLM population; the largest, 0.02 and 0.04, occur at the jumps, where single runs either ignite or not. Two rules complete the exposure-triggered cascade of § 2.1. If an agent’s exposure grows by more than one neighbor within a single round, it considers each new exposure level in turn and stops as soon as it decides to share. Sharing is irreversible: agents remain sharers for the rest of the cascade once they have shared the post. The effective response (§ 2.3) and the hidden-threshold picture of § 6.2 rely on both rules.
The degree-averaged response and the error observables
The degree-averaged response. The measured response of each degree is read as a function of the fraction of sharing contacts, and interpolated linearly between the points . The degree average is the plain average of these curves over all the degrees that occur in the network, each degree counting once, however many agents have it. On a homogeneous network it reduces to the single measured curve.
The averaging error (vertical axis of Figure 4A) is : the absolute difference between the final cascade size obtained with the degree-resolved response and with the degree average , averaged over the 25 seed fractions between 0.005 and 0.30 of the Monte Carlo simulations. We take the absolute value so that errors of opposite sign at different seeds do not cancel, and we use the same window of seeds for every post and network so that all cases are compared on the same footing.
The degree-dependence of the response (horizontal axis of Figure 4A) is
where is the probability that a randomly chosen neighbor of an agent has degree ( being the fraction of agents with degree ). We use the effective response because it is the one that enters the theory, and the weights because the propagated part of the error, which dominates near the tipping point, sums over degrees with these weights; § 6.7 shows what the weights would change.
Networks of controlled heterogeneity
To vary heterogeneity at fixed mean degree, we start from a regular graph in which every agent has contacts and rewire it step by step [44]. At each step an agent with more than contacts is chosen at random, and the link to its least-connected neighbor is moved to an agent chosen anywhere in the network, with probability proportional to its degree, among the agents with at least as many contacts as . Each step keeps the number of agents and of links, and hence the mean degree, fixed, while making the degrees more unequal: well-connected agents gain links at the expense of less-connected ones. We take snapshots after steps per agent, with 100 graphs of agents each, for (Figure 4B) and for and 12 (Figure S1).
We measure heterogeneity by : is the mean number of further contacts of an agent reached by following a link, the quantity that controls propagation in the theory, and at fixed mean degree carries the same information as . The errors are computed with the mean-field theory, which reproduced the Monte Carlo simulations in all twelve post–network cases of Figure 3 and lets us scan many networks without new simulations; they are averaged over 601 seed fractions in the window (§ 6.7).
Acknowledgments
F.B. acknowledges support from a University of Pennsylvania MindCORE fellowship. M.S. acknowledges support from Grant No. RYC2022-037932-I, funded by MCIN/AEI/10.13039/501100011033 and FSE+, and Grant No. CNS2023-144156, funded by MICIU/AEI/10.13039/501100011033 and the European Union NextGenerationEU/PRTR.
An AI assistant (Claude, Anthropic) was used to help edit the text and to write analysis and plotting code. The authors checked all of its output and take full responsibility for the content.
Supplementary Materials
The response to a single sharer decides whether a small seed spreads on a sparse network
On a sparse network, a cascade can grow from a small seed only if one sharing contact is often enough to make an agent with few contacts share (§ 2.2). This section gives that response for the four posts and evaluates Watts’ cascade condition with it.
Table S1 compares the four posts where they differ most: the response of an agent that follows only two or three accounts, one of which has shared.
Watts’ cascade condition [30] makes the consequence quantitative. is the mean number of further activations that a newly active agent triggers through neighbors that a single active contact suffices to switch on; above 1 a vanishing seed spreads globally. On ER , where degrees 2–4 carry 72% of , it is 0.79 for gun-right against 0.03–0.33 for the other posts. The value for gun-right is close enough to 1 that a seed of tips the network.
Each agent behaves as if it had a hidden threshold
The mean-field theory of § 2.3 requires a response that does not decrease with and a final state that does not depend on the order in which agents update. This section shows that our protocol, in which an agent is asked again at each new exposure, meets both requirements, because it is equivalent to a model in which each agent draws a fixed threshold once.
Let us consider an agent of degree . It is asked at in that order, each time with an independent YES probability , so the probability that it has said NO to all of the first queries is
and (eq. (2)). is non-decreasing, with and : it is the cumulative distribution of an integer random variable , which is the level at which the agent first says YES.
The same process can be described in a second way. Suppose the agent draws a threshold once, from the distribution , keeps it fixed, and shares as soon as reaches . At each level , its probability of sharing, given that it has not shared before, is then
as in the first description. The two descriptions are therefore the same process: is the probability of sharing at level given no earlier sharing, and is the distribution of the threshold. A smooth means thresholds spread over many levels; a step-like means a threshold common to all agents.
Two consequences follow. (a) is monotone in by construction, whatever looks like, including the non-monotonic we measure for some high-degree agents. Any irreversible cascade requires this property. The condition is part of the model: an agent none of whose neighbors has shared never sees the post and is never asked. (b) Once the thresholds are frozen, the final active set is the closure of the seed set under the rule and does not depend on the update order: synchronous rounds and an asynchronous queue give the same distribution of final states, and only the time course differs. The per-level querying of the protocol is what guarantees this: an agent whose jumps by three gets the same three chances it would have had if its three neighbors had shared in separate rounds.
Gleeson and Cahalane list three conditions for their tree calculation to describe the process on a random graph: the network is locally tree-like, each agent changes state at most once, and the final state does not depend on the update order. Under our protocol the second condition holds because sharing is irreversible, and the third is consequence (b). The first is an assumption about the network, and the theory makes a second one, implicit in writing it with the degree distributions and alone: the degrees of neighboring agents are uncorrelated. Both hold for configuration-model graphs, such as our Erdős–Rényi networks, whose degree assortativity, the correlation between the degrees at the two ends of a link, is zero on average. Our Barabási–Albert networks are slightly disassortative (degree assortativity on average), and the theory still reproduces the Monte Carlo simulations on them (Figure 3). Clustered networks violate the first assumption.
Mean-field theory of the cascade
The main text gives the fixed point of the theory, eq. (3), but describes only in words the two binomial sums of the response that enter it. This section writes them out, so that the final size can be computed from any response .
The theory we use to analytically describe the information cascades is the one of Gleeson and Cahalane [39]. An agent of degree , whose other neighbors are each active with probability , becomes active with probability
where the sum runs over the neighbors other than the one through which the agent is reached; an agent of degree whose neighbors are each active with probability is active with probability
With eq. (3) of the main text, these two sums give the final size of the cascade.
Comparing different degree averages
A degree-blind measurement on randomly chosen agents would weight each degree by how many agents have it, and one on the agents that a cascade reaches by how often the cascade reaches it, while our degree average is the plain average of the measured response curves over the degrees present in the network, each degree counting once (§ 2.3). This section shows that weighting the degrees does not remove the error.
With the measured response of degree read as a function of the fraction (Materials and Methods), the neighbor-weighted average is , where is the probability that a randomly chosen neighbor of an agent has degree and is the fraction of agents with degree . It is the response of the agents that a cascade reaches, with the weights of the mean-field equations. Weighting each degree instead by how many agents have it, with in place of , gives the agent-weighted average , the response of a randomly chosen agent. We compute the cascades of all three averages with the mean-field theory, which reproduces the Monte Carlo simulations for both the degree-resolved and the degree-averaged response on every network (Figure 3).
No degree average removes the error. The plain average leaves an error of at least 0.03 in eleven of the twelve post–network combinations; the exception is police-defunded on ER , with 0.013 (Table S2). An error of 0.03 means that the predicted final size is off by 3% of the agents, on average over the seed fractions. Both weighted averages lower the error on average, from 0.140 to 0.090 with the weights and to 0.076 with , but not in every combination: both raise it in the same three, gun-right on ER and police-defunded on ER and . The error is at least 0.029 with and at least 0.042 with , in every combination.
| post | network | plain average | agent-weighted average | neighbor-weighted average |
| migration-must-be-limited | ER | 0.126 | 0.029 | 0.048 |
| migration-must-be-limited | ER | 0.108 | 0.076 | 0.073 |
| migration-must-be-limited | BA | 0.190 | 0.075 | 0.076 |
| migration-must-be-stopped | ER | 0.075 | 0.042 | 0.044 |
| migration-must-be-stopped | ER | 0.072 | 0.046 | 0.069 |
| migration-must-be-stopped | BA | 0.363 | 0.092 | 0.042 |
| police-defunded | ER | 0.013 | 0.045 | 0.047 |
| police-defunded | ER | 0.068 | 0.136 | 0.071 |
| police-defunded | BA | 0.296 | 0.110 | 0.109 |
| gun-right | ER | 0.182 | 0.139 | 0.081 |
| gun-right | ER | 0.036 | 0.145 | 0.107 |
| gun-right | BA | 0.148 | 0.148 | 0.148 |
| mean over the twelve cases | 0.140 | 0.090 | 0.076 |
Table S2. Table S2: No degree average removes the averaging error. Averaging error mean: the average, over the 25 seed fractions between 0.005 and 0.30, of the absolute difference between the final size of the cascade with the degree-resolved response and with a degree average, as on the vertical axis of Figure 4A. All values are from the mean-field theory, for the plain average used in the main text, the agent-weighted average and the neighbor-weighted average .
Cascade dynamics on Barabási–Albert networks
The degree average fails most on BA (Figure 4A). This section shows how large its error is there and how the heterogeneity of the network causes it.
On BA networks the degree-resolved response keeps a finite ignition threshold that the mean-field reproduces, while the degree-averaged response ignites the network from a vanishingly small seed (tipping seed ). The degree-averaging error is therefore near-total below the tipping seed of the cascade with (, the gap between the two cascades in Figure 3D), and because is there it is outside the reach of the perturbative eqs. (11)–(12); the difference of the two mean-field solutions remains exact.
The error is this large because of how the degree average is built on a heterogeneous network. The average gives every degree that occurs in the network the same weight. On BA, 147 different degrees occur, but most agents (71%) have only 3 to 5 contacts: the degrees carry 94% of the weight in the average (), although only 11% of the agents have them. Since high-degree agents need a smaller fraction () of sharing contacts before they share (§ 2.2), the degree average ends up close to the response of a hub. For migration-must-be-limited, for example, crosses at , while the agents with 3 to 5 contacts, who make up most of the network, cross it between and . Applied to every agent, the degree average is therefore far too permissive, and the cascade with ignites the network from a vanishingly small seed: its tipping seed is at most 0.010, against 0.04–0.13 for the degree-resolved response (Table S3).
| post | network | tipping seed with | , fraction | , count | |
| migration-must-be-limited | ER | 0.214 | 0.126 | 0.498 | 0.192 |
| ER | 0.102 | 0.108 | 0.294 | 0.548 | |
| BA | 0.061 | 0.190 | 0.190 | 0.653 | |
| migration-must-be-stopped | ER | — | 0.075 | 0.601 | 0.076 |
| ER | 0.211 | 0.072 | 0.578 | 0.263 | |
| BA | 0.133 | 0.363 | 0.402 | 0.444 | |
| police-defunded | ER | 0.187 | 0.013 | 0.449 | 0.240 |
| ER | 0.135 | 0.068 | 0.388 | 0.454 | |
| BA | 0.099 | 0.296 | 0.296 | 0.551 | |
| gun-right | ER | 0.016 | 0.182 | 0.055 | 0.634 |
| ER | 0.058 | 0.036 | 0.184 | 0.658 | |
| BA | 0.044 | 0.148 | 0.148 | 0.698 |
Table S3. A response measured in one large group mispredicts the cascade. The response measured in the largest group (), given to every agent and read at the fraction or at the count , against the degree-resolved response and the degree-averaged response , from the mean-field theory. Tipping seed: the seed fraction at which the final size of the cascade with first reaches one half (—: not reached for ). Read at the fraction, the large-group response reaches one half at the smallest seed, , in every case; read at the count, it never does. Error: the average, over the 25 seed fractions between 0.005 and 0.30, of the absolute difference between the final size of the cascade with and with the approximation, as in Figure 4A and Table S2; the column repeats the plain-average column of Table S2.
A response measured in one large group
Studies of well-mixed populations measure one response in a large group in which every agent sees all the others, and use it for every agent (§ 2.3). This section tests whether such a response predicts the cascade on our networks. In our measurement an agent with contacts sees others, so our largest degree, , is the largest such group we have: an agent is told how many of the accounts it follows have shared. On a network, this response can be given to an agent with sharing contacts out of in two ways: read at the fraction, , the curve measured in the group of size evaluated at the fraction (interpolated linearly between the measured points); or read at the count, . We compute these cascades with the mean-field theory of § 2.3 only, without Monte Carlo simulations, because the theory reproduces them for both the degree-resolved and the degree-averaged response on every network (Figure 3), and it is least accurate near a tipping point, which neither reading of the large-group response has within the range of seeds. Table S3 gives the results.
Read at the fraction, the response ignites every network from the smallest seed, : an agent with few contacts shares as soon as one of them has shared, because one contact out of a few is a large fraction. Read at the count, it hardly spreads beyond the seed (the final size exceeds the seed fraction by at most 0.011, on BA, and by less than 0.001 on the ER networks): an agent with few contacts never reaches the count at which a member of a large group shares. The cascade with the degree-resolved response, which the LLM population follows (Figure 3), lies between these two extremes.
The error of the large-group response is at least as large as that of the degree-averaged response in 23 of the 24 comparisons. Read at the count, it is larger in all twelve cases, although only marginally (0.076 against 0.075) for migration-must-be-stopped on ER , whose cascade with does not reach one half either. Read at the fraction, it is larger in eight cases and equal in three, on BA, where both responses ignite the network from the smallest seed. The exception is gun-right on ER , with 0.055 against 0.182. There the cascade with also starts from a very small seed, because the measured response of this post to a single sharer is high at low degree (, Table S1); read at the fraction, the large-group response also makes an agent with few contacts share after a single sharer, so its agreement there is accidental.
The two errors differ in nature. The degree-averaged response is exact on a network in which every agent has the same degree, and its error grows with the heterogeneity of the network (Figure 4B). The large-group response gives every agent the response of a degree that almost no agent has (on BA, 0.04% of the agents have ; on the ER networks, no agent has more than 18 contacts), so it fails even when all agents have the same degree: on the regular networks from which the heterogeneity sweeps of § 2.4 start (, 7, 9 and 12), its error is between 0.19 and 0.65 read at the fraction, and between 0.16 and 0.66 read at the count.
Choices behind the degree dependence and the heterogeneity sweeps
Figure 4 rests on two choices that the main text states briefly: the weights in the degree dependence of the response, and the mean degree and the seed fractions of the heterogeneity sweeps. This section shows what other weights would change and why the sweeps are set up as they are.
The weights of the degree dependence. Eq. (5) weights each degree by because the propagated term of Eq. (4), which the susceptibility amplifies near the tipping point, sums over degrees with these weights. With the weights of the direct term instead, the degree dependence orders the error less well: the rank correlation between the error and the degree dependence over the twelve post–network cases drops from 0.92 to 0.84, and the error of gun-right no longer grows along its line in Figure 4A.
The heterogeneity sweeps. Figure 4B uses the mean degree , close to that of the ER and BA networks of Figure 4A, so that both panels compare networks of similar density. Figure S1 shows the sweeps at all four mean degrees. We average the error over 601 seed fractions in the window , instead of the 25 of the Monte Carlo simulations, because on these networks the transition is discontinuous: each of the 25 seeds that falls between the two tipping seeds adds about one twenty-fifth of the jump to the average, so a 25-seed average moves in steps of about 0.036 and can rise and fall even when the gap between the tipping seeds grows smoothly.
How the error depends on the mean degree
§ 2.4 draws no conclusion about how the error depends on the mean degree, because the trend depends on how the degree average is built. This section shows the trend for the plain and the two weighted averages (§ 6.4), and that the growth of the error with heterogeneity survives this choice.
The trend with mean degree depends on the average. In the heterogeneity sweeps of Figure S1, at the same heterogeneity, , and averaged over the four posts, the error grows with the mean degree for the plain average, from 0.082 at to 0.205 at , but falls for the neighbor-weighted one, from 0.056 to 0.028, and for the agent-weighted one, from 0.069 to 0.018. The growth of the error with heterogeneity survives the choice: its rank correlation with is positive in fifteen of the sixteen sweeps with the neighbor-weighted average, and in the exception, a sweep of migration-must-be-stopped, the error stays flat or falls slightly; with the agent-weighted average it is positive in eleven of the sixteen, all except the four sweeps of migration-must-be-stopped and the sweep of migration-must-be-limited at .
Expansion of the averaging error
§ 2.5 explains the errors of the degree average with a first-order expansion, eq. (4), and tests factorization with a single number for the degree dependence of the response. This section derives the expansion and the shift of the tipping seed, shows why no single number can order the errors, and
defines .
The error to first order. In the mean-field theory, eq. (3), the probability that a link leads to an active agent solves , and the final size is , with the smallest solution. The functions and are sums over degrees of the binomial averages and of the response, eqs. (8) and (S4), so they are linear in the response. Hence and , with and , where and . At the same , is the change in the probability that a link leads to an active agent, and the change in the probability that an agent is active. Let solve the fixed point with . To first order in , the fixed point with is . Inserting it into and keeping only first-order terms, cancels against , and what remains is , where is the derivative of with respect to . Solving for gives
The susceptibility says how strongly the fixed point responds to a change in : a link that leads to an extra active agent makes, on average, further links lead to active agents, these make more, and so on, and sums this chain. It diverges at the tipping point, where . Inserting into and keeping again the first-order terms gives , where is the derivative of with respect to . The error in the final size is therefore
The first term is the direct change: at the same probability that its neighbors are active, an agent with the response is active with a different probability than one with . The second term is the propagated change: the deviation of the response also changes the probability that a link leads to an active agent, which moves the fixed point by ; the susceptibility amplifies this change, and turns it into a change of the final size. Near the tipping point, where is large, the propagated term dominates. Writing and as sums over with eqs. (8) and (S4) gives eq. (4) of the main text. At small seed fractions, the sign of computed from the measured response matches the sign of the error in all twelve post–network combinations: is negative, and the degree average overpredicts the cascade, except for gun–right on ER .
The shift of the tipping seed. Where the cascade jumps, the final size is discontinuous at the tipping seed, and between the tipping seeds of the two cascades their final sizes differ by an amount that a first-order expansion cannot capture. There the useful observable is the shift of the tipping seed, . The tipping seed of the theory is located by two conditions, and : the curve touches the diagonal, so that the smallest fixed point disappears and the final size jumps (a saddle-node bifurcation). With instead of , the curve moves by , and the seed fraction at which it touches the diagonal moves by . Because the curve is tangent to the diagonal at the touching point , a small change of leaves the first condition unchanged to first order, so the change of the seed alone must compensate : , which gives
with the fixed point with at its tipping seed. The derivative is positive because a seed is active for certain, while any other agent is active only with probability , so a larger seed raises . In all our discontinuous cases the final size jumps at this seed from below one half to above it, so it is the tipping seed defined in § 2.1. Written out, it is again a weighted sum of , with . When , as in all our discontinuous cases, the tipping seed with is larger than the one with : the degree average underestimates how large a seed is needed to ignite the network.
Why no single number orders the errors. The sum of eq. (4) is a scalar product, , with . With the norm , the error can be written as , where is the angle between and . This splits the error into three factors. The first, , is the size of the response deviation. The second, , is the amplification by the network: it contains , and the susceptibility , and it diverges at the tipping point. The third, , is the alignment between and : it is large when the response deviates where the cascade samples it. A norm of sees only the first factor. The second depends on the network, not on the deviation of the response, and the third depends on the shape of , which changes from post to post. Therefore no single number that summarizes can order the errors of different posts, which is the failure of factorization of Section 2.5. By the Cauchy–Schwarz inequality, , and the bound is reached only when is aligned with .
The single number . For the factorization test of Section 2.5 we summarize the degree dependence of the response by a single number , the root-mean-square departure of the measured response from a pure fraction rule: is the average, over all measured , of , where is the average of the response curves over all measured degrees. We choose this because it depends only on the post and not on the network, which is what a factorization requires. On each network of the heterogeneity sweeps where the error exceeds 0.01, we divide the largest value of error/ among the four posts by the smallest; the median of this quotient over these networks is 2.1, the factor of about two of Section 2.5. A counterexample to factorization is migration-must-be-stopped: from on it has the largest error of the four posts at (Figure 4B), but a smaller than migration-must-be-limited (0.131 against 0.147).
Validation of the first-order expansion of the error against the Monte Carlo simulations
Section 2.5 states that the first-order expansion reproduces the measured error where the two cascades stay close. This section gives the evidence.
Figure S2 compares the expansion with the Monte Carlo simulation for each post: the pointwise error, eq. (11), on ER , where the transition is continuous (panel A), and the tipping-seed shift, eq. (12), on ER , where the transition is discontinuous and the shift is the well-posed observable (panel B). On BA networks the two cascades do not stay close, because is of order one, and the expansion does not apply (Section S5).
Figure S2. The first-order expansion against the Monte Carlo simulation, post by post. Error of the degree-averaged response against the seed fraction , for the four posts (columns). Black dots: Monte Carlo simulations; gray line: difference of the two mean-field solutions; dashed lines, colored by post as in Figure 4: first-order expansion, eq. (S6), computed from the solution with and the measured response alone. (A) ER , continuous transition, with the of the expansion against the Monte Carlo simulations from the smallest seed to the extremum of the error. For gun-right only the cascade with tips (dotted line), so the error is beyond the tipping seed and the expansion is shown only below it. (B) ER , discontinuous transition; gray band: the window between the two tipping seeds, whose shift Table S4 lists.
Continuous transition (ER ). In Figure S2A, eq. (4) is compared with the Monte Carlo error for the three posts for which both responses stay continuous, migration-must-be-limited, migration-must-be-stopped and police-defunded. We compute from the smallest seed up to the extremum of the measured error, because this range is where the expansion is expected to hold; beyond the extremum the susceptibility of eq. (4), which amplifies the response deviation near the tipping point, grows and the leading-order expansion degrades. It follows the error through and past its minimum for police-defunded (). For migration-must-be-stopped it reproduces the shape of the error but overestimates its size, by about 14% at the deepest point (), and for migration-must-be-limited by about 34% (). The overestimate grows with the depth of the error, which is 0.03 for police-defunded, 0.13 for migration-must-be-stopped and 0.19 for migration-must-be-limited, as one expects from an expansion that keeps only the first order in , while the difference of the two mean-field solutions follows the Monte Carlo error for all three. For gun-right the error is beyond the tipping seed of the cascade with , where the expansion does not apply; once the cascade with has also spread, at larger seeds, the difference of the two mean-field solutions tracks the Monte Carlo error closely.
Discontinuous transition (ER ). Table S4 lists the tipping-seed shift predicted by the closed form eq. (12) against the Monte Carlo shift. The prediction is accurate for all four posts; it falls short only for migration-must-be-limited, the post with the largest shift (leading-order truncation). The theory and the Monte Carlo simulations use the same tipping seed, the seed fraction at which the final size crosses one half (§ 2.1). The theory locates it where its smallest fixed point disappears and the final size jumps through one half (Section S9); the Monte Carlo simulations locate it where the average final size crosses one half.
| post (ER ) | expansion, eq. (12) | Monte Carlo | mean-field difference, all orders |
| migration-must-be-stopped | +0.027 | +0.026 | +0.026 |
| police-defunded | +0.023 | +0.023 | +0.022 |
| gun-right | +0.013 | +0.013 | +0.013 |
| migration-must-be-limited | +0.035 | +0.039 | +0.040 |
Table S4. The expansion predicts the shift of the tipping seed. Shift of the tipping seed on ER , from the first-order expansion, eq. (12); from the Monte Carlo simulations; and from the difference of the two mean-field solutions, to all orders in . The expansion matches the Monte Carlo simulations for every post, within 10%.
References
- [1]Yingxuan Yang, Mulei Ma, Yuxuan Huang, Huacan Chai, Chenyu Gong, Haoran Geng, Yuanjian Zhou, Ying Wen, Meng Fang, Muhao Chen, Shangding Gu, Ming Jin, Costas Spanos, Yang Yang, Pieter Abbeel, Dawn Song, Weinan Zhang, and Jun Wang. Agentic web: Weaving the next web with AI agents. arXiv preprint arXiv:2507.21206, 2025.arxiv.org/abs/2507.21206
- [2]Yuanchun Li, Hao Wen, Weijun Wang, Xiangyu Li, Yizhen Yuan, Guohong Liu, et al. Personal LLM agents: Insights and survey about the capability, efficiency and security. arXiv preprint arXiv:2401.05459, 2024.arxiv.org/abs/2401.05459
- [3]Henry C. W. Price, H. AlMuhanna, P. M. Bassani, M. Ho, and T. S. Evans. Let there be claws: An early social network analysis of AI agents on Moltbook. arXiv preprint arXiv:2602.20044, 2026.arxiv.org/abs/2602.20044
- [4]Lingyao Li, Renkai Ma, Chen Chen, Zhicong Lu, and Yongfeng Zhang. The rise of AI agent communities: Large-scale analysis of discourse and interaction on Moltbook. arXiv preprint arXiv:2602.12634, 2026.arxiv.org/abs/2602.12634
- [5]Ryan Greenblatt, Ajeya Cotra, and Hjalmar Wijk. Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI/Hugging Face hacking incident. METR and Redwood Research, https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/, 2026.
- [6]Levin Brinkmann, Fabian Baumann, Jean-François Bonnefon, Maxime Derex, Thomas F. Müller, Anne-Marie Nussberger, Agnieszka Czaplicka, Alberto Acerbi, Thomas L. Griffiths, Joseph Henrich, et al. Machine culture. Nature Human Behaviour, 7(11):1855–1868, 2023.
- [7]Claudio Castellano, Santo Fortunato, and Vittorio Loreto. Statistical physics of social dynamics. Reviews of Modern Physics, 81(2):591–646, 2009.DOI
- [8]Damon Centola and Andrea Baronchelli. The spontaneous emergence of conventions: An experimental study of cultural evolution. Proceedings of the National Academy of Sciences, 112(7):1989–1994, 2015.DOI
- [9]Ariel Flint Ashery, Luca Maria Aiello, and Andrea Baronchelli. Emergent social conventions and collective bias in LLM populations. Science Advances, 11(20):eadu9368, 2025.DOI
- [10]Fabian Baumann, Philipp Lorenz-Spreen, Igor M. Sokolov, and Michele Starnini. Modeling echo chambers and polarization dynamics in social networks. Physical Review Letters, 124(4):048301, 2020.DOI
- [11]Jinghua Piao, Zhihong Lu, Chen Gao, Fengli Xu, Qinghua Hu, Fernando P. Santos, Yong Li, and James Evans. Emergence of polarization in networks of large language model agents. Nature Communications, 17:11463, 2026.arxiv.org/abs/2501.05171
- [12]Soroush Vosoughi, Deb Roy, and Sinan Aral. The spread of true and false news online. Science, 359(6380):1146–1151, 2018.DOI
- [13]Ziyi Yang, Zaibin Zhang, Zirui Zheng, Yuxian Jiang, Ziyue Gan, Zhiyu Wang, Zijian Ling, Jinsong Chen, Martz Ma, Bowen Dong, Prateek Gupta, Shuyue Hu, Zhenfei Yin, Guohao Li, Xu Jia, Lijun Wang, Bernard Ghanem, Huchuan Lu, Chaochao Lu, Wanli Ouyang, Yu Qiao, Philip Torr, and Jing Shao. OASIS: Open agent social interaction simulations with one million agents. arXiv preprint arXiv:2411.11581, 2024.arxiv.org/abs/2411.11581
- [14]Genglin Liu, Vivian Le, Salman Rahman, Elisa Kreiss, Marzyeh Ghassemi, and Saadia Gabriel. MOSAIC: Modeling social AI for content dissemination and regulation in multi-agent simulations. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2025.arxiv.org/abs/2504.07830
- [15]Lan Zhang, Yuxuan Hu, Weihua Li, Quan Bai, and Parma Nand. LLM-AIDSim: LLM-enhanced agent-based influence diffusion simulation in social networks. Systems, 13(1):29, 2025.DOI
- [16]Lewis Hammond, Alan Chan, Jesse Clifton, Jason Hoelscher-Obermaier, Akbir Khan, Euan McLean, Chandler Smith, Wolfram Barfuss, Jakob Foerster, Tomáš Gavenčiak, et al. Multi-agent risks from advanced AI. arXiv preprint arXiv:2502.14143, 2025.arxiv.org/abs/2502.14143
- [17]Daniel Thilo Schroeder, Meeyoung Cha, Andrea Baronchelli, Nick Bostrom, Nicholas A. Christakis, David Garcia, et al. How malicious AI swarms can threaten democracy. Science, 391(6783):354–357, 2026.
- [18]Ariel Flint, Luca Maria Aiello, Sara M. Constantino, Romualdo Pastor-Satorras, and Andrea Baronchelli. Indirect tipping: a social attack surface in AI agent populations. arXiv preprint arXiv:2609.25194, 2026.arxiv.org/abs/2609.25194
- [19]Thomas C. Schelling. Micromotives and Macrobehavior. W. W. Norton, New York, 1978.
- [20]Mark Granovetter. Threshold models of collective behavior. American Journal of Sociology, 83(6):1420–1443, 1978.DOI
- [21]Michele Starnini, Fabian Baumann, Tobias Galla, David Garcia, Gerardo Iñiguez, Márton Karsai, Jan Lorenz, and Katarzyna Sznajd-Weron. Opinion dynamics: Statistical physics and beyond. Reviews of Modern Physics, 98(3):035004, 2026.arxiv.org/abs/2507.11521
- [22]Damon Centola, Joshua Becker, Devon Brackbill, and Andrea Baronchelli. Experimental evidence for tipping points in social convention. Science, 360(6393):1116–1119, 2018.DOI
- [23]James Andreoni, Nikos Nikiforakis, and Simon Siegenthaler. Predicting social tipping and norm change in controlled experiments. Proceedings of the National Academy of Sciences, 118(16):e2014893118, 2021.DOI
- [24]Daniel M. Romero, Brendan Meeder, and Jon Kleinberg. Differences in the mechanics of information diffusion across topics: idioms, political hashtags, and complex contagion on Twitter. In Proceedings of the 20th International Conference on World Wide Web, pages 695–704, 2011.DOI
- [25]Steven D. Levitt and John A. List. What do laboratory experiments measuring social preferences reveal about the real world? Journal of Economic Perspectives, 21(2):153–174, 2007.
- [26]Giordano De Marzo, Claudio Castellano, and David Garcia. Ai agents can coordinate via majority-following beyond human scale. Science Advances, 12(33):eaea6091, 2026.DOI
- [27]Ariel Flint, Luca Maria Aiello, Romualdo Pastor-Satorras, and Andrea Baronchelli. Group size effects and collective misalignment in LLM multi-agent systems. Proceedings of the National Academy of Sciences, 123(34):e2531697123, 2026.DOI
- [28]B. El, J. Paeng, F. Dinc, S. Su, M. Erdogan, A. Pappu, H. Ye, W. Zhao, S. Ganguli, and J. Zou. Physics of agents: Statistical mechanics predicts collective behavior of AI agents. arXiv preprint arXiv:2608.16578, 2026.arxiv.org/abs/2608.16578
- [29]Roberto Passaro, Ariel Flint, Romualdo Pastor-Satorras, Andrea Baronchelli, and Luca Maria Aiello. Network topology shapes consensus and vulnerability in LLM populations. arXiv preprint arXiv:2610.31165, 2026.arxiv.org/abs/2610.31165
- [30]Duncan J. Watts. A simple model of global cascades on random networks. Proceedings of the National Academy of Sciences, 99(9):5766–5771, 2002.
- [31]Peter Sheridan Dodds and Duncan J Watts. Universal behavior in a generalized model of contagion. Physical review letters, 92(21):218701, 2004.DOI
- [32]Damon Centola and Michael Macy. Complex contagions and the weakness of long ties. American Journal of Sociology, 113(3):702–734, 2007.DOI
- [33]Joon Sung Park, Joseph O’Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, pages 1–22, 2023.arxiv.org/abs/2304.03442
- [34]Lisa P. Argyle, Ethan C. Busby, Nancy Fulda, Joshua R. Gubler, Christopher Rytting, and David Wingate. Out of one, many: Using language models to simulate human samples. Political Analysis, 31(3):337–351, 2023.DOI
- [35]Elif Akata, Lion Schulz, Julian Coda-Forno, Seong Joon Oh, Matthias Bethge, and Eric Schulz. Playing repeated games with large language models. Nature Human Behaviour, 9(7):1380–1390, 2025.DOI
- [36]F. Nøhr Laustsen, M. Haahr Petersen, V. Popa, A. Flint, R. Pastor-Satorras, A. Baronchelli, and L. M. Aiello. Peer influence across heterogeneous AI models. arXiv preprint arXiv:2610.03095, 2026.arxiv.org/abs/2610.03095
- [37]Shibani Santurkar, Esin Durmus, Faisal LadhaK, Cinoo Lee, Percy Liang, and Tatsunori Hashimoto. Whose opinions do language models reflect? In Proceedings of the 40th International Conference on Machine Learning, PMLR, volume 202, pages 29971–30004, 2023.arxiv.org/abs/2303.17548
- [38]Yun-Shiuan Chuang, Agam Goyal, Nikunj Harlalka, Siddharth Suresh, Robert Hawkins, Sijia Yang, Dhavan Shah, Junjie Hu, and Timothy Rogers. Simulating opinion dynamics with networks of LLM-based agents. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 3326–3346, 2024.arxiv.org/abs/2311.09618
- [39]James P Gleeson and Diarmuid J Cahalane. Seed size strongly affects cascades on random networks. Physical Review E—Statistical, Nonlinear, and Soft Matter Physics, 75(5):056103, 2007.pubmed.ncbi.nlm.nih.gov/17677129
- [40]James P. Gleeson. Binary-state dynamics on complex networks: Pair approximation and beyond. Physical Review X, 3(2):021004, 2013.
- [41]William J. Brady, Julian A. Wills, John T. Jost, Joshua A. Tucker, and Jay J. Van Bavel. Emotion shapes the diffusion of moralized content in social networks. Proceedings of the National Academy of Sciences, 114(28):7313–7318, 2017.DOI
- [42]James P Gleeson. Cascades on correlated and modular random networks. Physical Review E—Statistical, Nonlinear, and Soft Matter Physics, 77(4):046117, 2008.pubmed.ncbi.nlm.nih.gov/18517700
- [43]Rebekka Burkholz and Frank Schweitzer. Correlations between thresholds and degrees: An analytic approach to model attacks and failure cascades. Physical Review E, 98(2):022306, 2018.DOI
- [44]Jaume Ojer, Michele Starnini, and Romualdo Pastor-Satorras. Social network heterogeneity promotes depolarization of multidimensional correlated opinions. Physical Review Research, 7(1):013207, 2025.DOI
- [45]Márton Karsai, Gerardo Iñiguez, Riivo Kikas, Kimmo Kaski, and János Kertész. Local cascades induced global contagion: How heterogeneous thresholds, exogenous effects, and unconcerned behaviour govern online adoption spreading. Scientific Reports, 6:27178, 2016.DOI
- [46]Nathan O. Hodas and Kristina Lerman. How visibility and divided attention constrain social contagion. In Proceedings of the 2012 ASE/IEEE International Conference on Social Computing (SocialCom), pages 249–257, 2012.DOI
- [47]Manuel Gomez-Rodriguez, Krishna P. Gummadi, and Bernhard Schölkopf. Quantifying information overload in social media and its impact on social contagions. In Proceedings of the International AAAI Conference on Web and Social Media, volume 8, pages 170–179, 2014.arxiv.org/abs/1403.6838
- [48]Gueorgi Kossinets and Duncan J. Watts. Empirical analysis of an evolving social network. Science, 311(5757):88–90, 2006.pubmed.ncbi.nlm.nih.gov/16400149
- [49]Thilo Gross and Bernd Blasius. Adaptive coevolutionary networks: a review. Journal of the Royal Society Interface, 5(20):259–271, 2008.DOI
- [50]Adam Hackett, Sergey Melnik, and James P. Gleeson. Cascades on a class of clustered random networks. Physical Review E, 83(5):056107, 2011.pubmed.ncbi.nlm.nih.gov/21728605
- [51]Paul Erdős and Alfréd Rényi. On random graphs I. Publicationes Mathematicae Debrecen, 6:290–297, 1959.
- [52]Albert-László Barabási and Réka Albert. Emergence of scaling in random networks. Science, 286(5439):509–512, 1999.



