Giordano De Marzo

Large language models (LLMs) are increasingly deployed in collaborative tasks forming “AI agent societies” where agents interact and influence one another. Whether such groups can spontaneously coordinate without external influence, a hallmark of self-organized regulation in human societies, remains an open question. Here, we use principles from complexity and behavioral science to investigate coordination in AI agent groups through majority-following, a fundamental mechanism for spontaneous consensus formation. Using binary opinion dynamics experiments across multiple LLM architectures and group sizes, we find that agents exhibit majority-following characterized by a universal functional form with a single parameter, the “majority force.” This majority force diminishes as group size increases, leading to a critical size beyond which coordination becomes unattainable. The critical group size grows rapidly with model capabilities and, for advanced LLMs, exceeds 1000 agents, larger than typical human informal groups. Our findings have implications for designing collaborative AI systems where coordination could be beneficial or pose safety threats.

As AI agents increasingly operate in multi-agent settings, alongside other models and humans, understanding how they respond to the opinions they find in their prompt becomes critical for predicting the outcome of these interactions. This study examines conformity-like behavior, the tendency to align with group opinions, in large multimodal language models functioning as AI agents. By adapting classic visual experiments from social psychology, we investigate how these agents respond to social signals in the prompt. Our experiments reveal that AI agents exhibit a systematic conformity bias consistent with Social Impact Theory. Their responses vary with majority size, group unanimity, task difficulty, and source characteristics, including perceived authority and shared group membership. Critically, AI agents achieving near-perfect performance in isolation become highly susceptible to manipulation through social influence. This vulnerability persists across model scales: while larger models show reduced conformity bias on simple tasks, they remain vulnerable when the task is difficult. These findings reveal security vulnerabilities in AI agent decision-making that could enable malicious manipulation, misinformation campaigns, and bias propagation in interactions with other models and with humans, highlighting the urgent need for safeguards in AI deployments.

Systems driven by innovation, a pivotal force in human society, present various intriguing statistical regularities, from the Heaps' law to logarithmic scaling or somewhat different patterns for the innovation rates. The urn model with triggering (UMT) has been instrumental in modeling these innovation dynamics. Yet a generalization is needed to capture the richer empirical phenomenology. Here we introduce a time-dependent urn model with triggering (TUMT), a generalization of the UMT that crucially integrates time-dependent parameters for reinforcement and triggering to offer a broader framework for modeling innovation in nonstationary systems. Through analytical computation and numerical simulations, we show that the TUMT reconciles various behaviors observed in a broad spectrum of systems from patenting activity to the analysis of gene mutations. We highlight how the TUMT features a “critical” region where both Heaps' and Zipf's laws coexist, for which we compute the exponents. Published by the American Physical Society 2025

As often happens in science, tools, and methods originally developed in one field can unexpectedly become useful in others. This paper explores the formalism of Economic Fitness Complexity (EFC), initially designed to predict and explain the economic trajectories of countries, cities, and regions, which has also proven applicable in diverse contexts such as ecology and chess openings. The success of EFC is attributed to its ability to indirectly assess hidden capabilities within a system. However, existing EFC algorithms are constrained to bipartite graphs, becoming inapplicable even with minor deviations in the bipartite structure. This paper introduces an extension of EFC and its cousin Economic Complexity Index that applies to any graph, thereby overcoming the bipartite constraint. This extension introduces fitness centrality, a novel centrality measure that can be used for assessing node vulnerability. By broadening the scope of economic complexity analysis to diverse network structures, this work expands the applicability and robustness of EFC in complexity science.

We conducted an analysis of the velocity field of dwarf galaxies in the sample collected by the Local Irregulars That Trace Luminosity Extremes and The H I Nearby Galaxy Survey (LITTLE THINGS) and focused on deriving 2D velocity maps that encompass the transverse and radial velocity fields. Within the range of radial distances in which velocity anisotropies are sufficiently small for the disk to be considered supported by rotation and where the warped geometry of the disk can be neglected, we reconstructed the rotation curve while taking the effect of the asymmetric drift into account. To fit the rotation curves, we employed the standard halo model and the dark matter disk (DMD) model, which assumes that dark matter is primarily confined to the galactic disks and can be traced by the distribution of H I. Interestingly, our analysis revealed that the fits from the DMD model are statistically comparable to those obtained using the standard halo model, but the inferred masses of the galaxies in the DMD model are approximately 10 to 100 times lower than the masses inferred in the standard halo model. In the DMD model, the inner slope of the rotation curve is directly related to a linear combination of the surface density profiles of the stellar and gas components, which generally exhibit a flat core. Consequently, the observation of a linear relation between the rotation curve and the radius in the central disk regions is consistent with the framework of the DMD model.

Studying how we explore the world in search of novelties is key to understand the mechanisms that can lead to new discoveries. Previous studies analyzed novelties in various exploration processes, defining them as the first appearance of an element. However, novelties can also be generated by combining what is already known. We hence define higher-order novelties as the first time two or more elements appear together, and we introduce higher-order Heaps’ exponents as a way to characterize their pace of discovery. Through extensive analysis of real-world data, we find that processes with the same pace of discovery, as measured by the standard Heaps’ exponent, can instead differ at higher orders. We then propose to model an exploration process as a random walk on a network in which the possible connections between elements evolve in time. The model reproduces the empirical properties of higher-order novelties, revealing how the network we explore changes over time along with the exploration process.

Recommender systems are vital for shaping user online experiences. While some believe they may limit new content exploration and promote opinion polarization, a systematic analysis is still lacking. In this paper we present a model that explores the influence of recommender systems on novel content discovery. Surprisingly, analytical and numerical findings reveal that these techniques can enhance novelty discovery rates. Also, distinct algorithms with similar discovery rates yield different outcomes, with the matrix factorization algorithm producing opinion polarization. Our approach shed light on the interplay between algorithmic recommendations and novelties discovery, offering a framework to enhance recommendation techniques beyond accuracy metrics.

A central role in shaping the experience of users online is played by recommendation algorithms. On the one hand they help retrieving content that best suits users taste, but on the other hand they may give rise to the so-called “filter bubble” effect, favoring the rise of polarization. In the present paper we study how a user-user collaborative-filtering algorithm affects the behavior of a group of agents repeatedly exposed to it. By means of analytical and numerical techniques we show how the system stationary state depends on the strength of the similarity and popularity biases, quantifying respectively the weight given to the most similar users and to the best rated items. In particular, we derive a phase diagram of the model, where we observe three distinct phases: disorder, consensus, and polarization. In the last users spontaneously split into different groups, each focused on a single item. We identify, at the boundary between the disorder and polarization phases, a region where recommendations are nontrivially personalized without leading to filter bubbles. Finally, we show that our model well reproduces the behavior of users on the online music platform last.fm. This analysis paves the way to a systematic analysis of recommendation algorithms by means of statistical physics methods and opens the possibility of devising less polarizing recommendation algorithms.

Chess is a centuries-old game that continues to be widely played worldwide. Opening Theory is one of the pillars of chess and requires years of study to be mastered. In this paper, we use the games played in an online chess platform to exploit the “wisdom of the crowd” and answer questions traditionally tackled only by chess experts. We first define a relatedness network of chess openings that quantifies how similar two openings are to play. Using this network, we identify communities of nodes corresponding to the most common opening choices and their mutual relationships. Furthermore, we demonstrate how the relatedness network can be used to forecast future openings players will start to play, with back-tested predictions outperforming a random predictor. We then apply the Economic Fitness and Complexity algorithm to measure the difficulty of openings and players’ skill levels. Our study not only provides a new perspective on chess analysis but also opens the possibility of suggesting personalized opening recommendations using complex network theory.