Source-linked AI summary
The Spontaneous Emergence of Conventions: An Experimental Study of Cultural Evolution
Damon Centola, Andrea Baronchelli
TL;DR
The paper examines whether shared conventions can emerge from local interactions without institutions that provide global incentives or information, a question limited by prior large-population evidence. Using Web-based naming-game experiments across network structures and scales, it finds that connectivity directs convention formation, with homogeneous mixing producing rapid universal adoption despite participants’ lack of population-level information.
Problem
The paper asks whether local interactions can produce universally adopted conventions without pre-existing institutions, addressing limited empirical insight into this process in larger populations.
Method
The study uses Web-based naming-game experiments that vary social network structure and examine convention formation across multiple population sizes.
Results
Increased network connectivity accelerated convergence to a global convention, and every respondent in homogeneously mixing networks knew the norm despite no information about the population size.
Takeaways & Limitations
Social conventions can spontaneously evolve from local coordination, while network structure directs whether dynamics produce competing local norms or global adoption.
Takeaways & Limitations
Small-group experiments may differ qualitatively from larger-group dynamics, limiting their ability to demonstrate convention formation in larger populations.
Abstract
from arXiv · showhide
How do shared conventions emerge in complex decentralized social systems? This question engages fields as diverse as linguistics, sociology and cognitive science. Previous empirical attempts to solve this puzzle all presuppose that formal or informal institutions, such as incentives for global agreement, coordinated leadership, or aggregated information about the population, are needed to facilitate a solution. Evolutionary theories of social conventions, by contrast, hypothesize that such institutions are not necessary in order for social conventions to form. However, empirical tests of this hypothesis have been hindered by the difficulties of evaluating the real-time creation of new collective behaviors in large decentralized populations. Here, we present experimental results - replicated at several scales - that demonstrate the spontaneous creation of universally adopted social conventions, and show how simple changes in a population's network structure can direct the dynamics of norm formation, driving human populations with no ambition for large scale coordination to rapidly evolve shared social conventions.
Introduction
The paper asks whether local interactions can produce globally shared conventions without institutions that coordinate or inform the population. It addresses empirical gaps involving scale, innovation, network complexity, and confounding institutional mechanisms.
- The central question is whether local interactions can spontaneously self-organize into universally adopted conventions without pre-existing social institutions.
- Social conventions are widespread within economic communities yet vary broadly across communities.
- Existing theories emphasize centralized authority, collective-agreement incentives, leadership, or aggregated information as mechanisms for global coordination.
- Prior evidence is limited because dynamics in dyads and small groups can differ qualitatively from those in larger populations.
- Earlier studies often used fixed alternatives with known payoffs or assumed rewards and information for universal coordination, unlike natural convention formation with continual innovation.
- The study uses Web-based experiments to test whether network structure shapes spontaneous linguistic convention formation when options have no prior value or advantage.
Theoretical model
The theoretical framework extends language-game models in which pairwise interaction, memory, and network connectivity shape convention formation. It compares three network configurations to test how topology affects collective dynamics.
- The language-game model has agents accrue memories of past pairwise plays and use them to guess words for subsequent partners.
- The model predicts that actors’ network connectivity can influence the collective dynamics of convention formation.
- The framework allows outcomes ranging from competing regional norms that inhibit global coordination to rapidly growing universally shared conventions.
- The experiments evaluate spatial lattices, random graphs, and homogeneously mixing populations.
- Formal results place alternative network configurations within the dynamical range exhibited by these three topologies.
Experimental Design
Participants played repeated naming games on assigned social networks, interacting with randomly selected neighbors and receiving rewards or penalties based on coordination. They saw only their own and partners’ choices after each round.
- Each trial assigned participants to positions in a specified social network and ran for a pre-specified number of rounds.
- In each round, two network neighbors were randomly selected to simultaneously name the same pictured object.
- Successful coordination produced a payment, whereas failure incurred a penalty.
- After each round, participants could see only their own and their partner’s choices.
- The object and naming task remained constant, while participants lacked information about population size and number of neighbors.
Results
Network structure decisively shaped convention formation: spatial and random networks sustained competing local conventions, whereas homogeneous mixing produced rapid global coordination. These patterns replicated across population sizes, including N=96.
- Spatial networks: Spatial networks developed competing local conventions whose entrenched boundaries impeded global coordination.
- Spatial and random networks: In spatial and random networks, the dominant convention averaged 33% of the population across trials.
- Homogeneous mixing: Homogeneously mixing populations initially had lower success because repeated partner interactions did not form entrenched behavioral neighborhoods.
- Scaling: The topology-dependent dynamics replicated at N=24 and N=48, while homogeneous mixing also produced global coordination at N=96 on a timescale comparable to N=24.
Discussion
Controlled experiments found spontaneous convention formation across network sizes and connectivity conditions, without participants receiving global network information. The results also indicate that connectivity shapes convergence dynamics and that findings remain robust when initial name preferences are controlled.
- Experimental controls: Across all network conditions and sizes, there were no significant differences in subjects’ reported network information.The design kept informational resources identical across conditions.
- Network-dependent dynamics: In homogeneously mixing networks, every respondent knew the norm even though none knew how many people were using it.This pattern differed from the other network conditions.
- Robustness checks: Randomized ten-name-list experiments were indistinguishable from the main results, ruling out convergence driven by pre-existing name popularity.The name order was randomized separately for each participant to avoid implicit ranking effects.
- Robustness checks: The emergent name ecology contained more suggested names than participants, sometimes by more than a factor of 2, indicating no initially preferred options limited choice.This diversity pattern was observed in every trial.
- Discussion: The experiments show that large human populations can spontaneously evolve conventions without institutional coordination mechanisms, while network connectivity plays a causal role in establishing shared norms.Increased connectivity accelerated convergence to a global norm, and the findings contrast with earlier work focused on only two payoff-differentiated norms.
- Experimental design: Participants were randomly assigned to social networks and repeatedly named the same object with randomly selected neighbors, receiving rewards for coordination.They lacked information about population size, network degree, and partner identities.
Figures
The figures compare convention formation across spatial, random, and homogeneous-mixing networks, showing distinct patterns of local coordination, competition, and global dominance.
- Figure 1 compares evolving conventions across spatial, random, and homogeneous-mixing networks at N=24.
- Spatial and random networks formed competing local groups, whereas homogeneous mixing produced rapid global coordination on one convention.
- By round 16, one name dominated homogeneous mixing while spatial and random networks retained competing groups without a winner.
- Average dominant-convention size was 30% in spatial networks, 33% in random networks, and 96% in homogeneous-mixing populations.
Homogeneous mixing populations
In homogeneously mixing populations, small fluctuations break symmetry between conventions and are amplified until one convention becomes globally dominant.
- The increasingly favored convention progressively eliminates competitors, producing consensus in O(N^0.5) Rounds Played.
- For population size N, consensus takes approximately O(N^1.5) microscopic interactions.
- Stochastic fluctuations break initial symmetry between conventions, making one more popular and allowing interaction dynamics to amplify its advantage.
- Pure imitation models differ because fluctuations can reverse the advantage of competing states before consensus is reached.
Networked populations
Network topology shapes convention dynamics: repeated local interactions promote regional coordination, while shortcuts and global mixing accelerate convergence.
- Network nodes represent agents, links specify allowed communication channels, and network properties significantly affect overall model dynamics.
- Lattices: On low-dimensional lattices, repeated neighbor interactions favor local consensus and generate clusters that later compete through interface fluctuations.
- Lattices: Lattice convergence time scales as O(N2/d), where d ≤ 4 is spatial dimensionality.
- Small-world networks: In small-world networks, shortcuts restore fast mean-field-like convergence while finite connectivity preserves neighbor coordination.
- Small-world networks: For times beyond tcross = O(N/p2), small-world dynamics become shortcut-dominated and convergence scales as N1/2.
- Complex networks: Convergence-time scaling is robust to average degree, clustering, and degree-distribution details across complex network topologies.
3. Robustness of the model
The theoretical model closely matches individual and collective behavior, while simulations extend experimentally constrained observations to larger populations and timescales.
- Memory and repeated exposure to a successful term are essential: both agents must have heard the convention previously to coordinate successfully.
- Participants’ choices agreed with the theoretical model 95% of the time.
- Simulations with 95% model-rule behavior and 5% random entries produced collective dynamics indistinguishable from the theoretical model.
- The model reproduced experimental Name Game results and provided a good fit to measured user behavior.
- Experimental constraints limited accessible parameters, especially experiment duration and population size, motivating larger-scale model predictions.
SI References
The supplementary references draw on nonequilibrium lattice models, naming-game research, and studies of convention spreading in open-ended systems.
- The references include work on nonequilibrium phase transitions in lattice models.
- They cite empirical and theoretical studies of shared vocabularies and naming games in multi-agent and networked systems.
- The list also includes research on effective surface tension, reputation, and convention spreading in open-ended systems.
SI text Figures
The supplementary materials describe randomized experimental assignment, repeated name-matching interactions, and measures of success, naming diversity, beliefs, and convergence across network structures. Simulations and experiments show that network topology shapes the speed and pathway of convention formation, from prolonged spatial competition to rapid global convergence under homogeneous mixing.
- Experimental setup: Subjects were randomly assigned to experimental conditions and then randomly assigned to nodes within the selected social network.
- Experimental setup: Each round asks a subject to enter a name, observe the partner’s choice, and receive a reward only when the choices match.
- Measures: A matching interaction is coded as success with value 1, whereas a mismatch is coded as failure with value 0.
- Measures: The Player Success Rate averages binary outcomes over N/2 individual rounds, corresponding to one Round Played.
- Naming dynamics: Alternative names exceeded population size, while the reported number of words counted distinct spellings in active circulation and ignored case differences.
- Network effects: Random networks reached convergence by 180 rounds, while homogeneous mixing produced sharp symmetry breaking and a global convention on experimentally observable timescales.
- Network effects: Spatial networks formed local coordination rapidly but required more than 1000 rounds for final consensus because emergent groups competed locally.