All seminars are on Mondays at 8:30 am PT / 11:30 am ET / 4:30 pm London / 5:30 pm Berlin.
Date: Sep 9, 2024
Speaker: Hannah Rubin
Title: Diversity and homophily in group inquiry
Abstract: Diversity, broadly construed, is important to successful inquiry within a community. As one instance, diversity according to social identity has been argued to be important to inquiry, resulting in many arguments that we ought to promote demographic diversity because of the ensuing gains in effective inquiry or performance of groups. However, the potential benefits to diversity do not always materialize and efforts to promote diversity can backfire, both impeding inquiry and resulting in increased inequity. We argue that one reason for the gap between potential and realized benefits of demographic diversity is limitations to the inferences we can draw based on experimental results in which small groups are shown to benefit from decreased group-based conformity. We use formal models to show that factors that may be beneficial to group inquiry in restricted experimental settings (e.g., where everyone is talking to everyone) may be detrimental when we consider them in the context of how larger groups interact and share information. We argue that this has consequences for both how we attempt to justify and how we attempt to implement diversity initiatives. (This is joint work with Sina Fazelpour.)
Date: Sep 23, 2024
Speaker: Alice Huang
Title: Landscapes and Bandits: A Unified Model of Functional and Demographic Diversity
Abstract: Two types of formal models - landscape search tasks and two-armed bandit models - are often used to study the effects that various social factors have on epistemic performance. I argue that they can be understood within a single framework. In this unified framework, I develop a model that may be used to understand the effects of functional and demographic diversity and their interaction. Using the unified model, I find that the benefit of demographic diversity is most pronounced in a functionally homogeneous group, and decreases with the increase of functional diversity.
Date: Oct 7, 2024
Speaker: Felix Kopecky
Title: Diversity in Dialectical Structures
Abstract: The theory of dialectical structures offers a rich if lesser-known approach to computational social epistemology. In models built on the theory, agents put forward arguments and adjust their beliefs to arguments introduced by others, thereby constructing argument maps, or “dialectical structures”, of different shapes and sizes. In this talk, I will briefly go back to the theory's origins and review some earlier computational results regarding disagreement and belief polarisation. I will then present results on a model variant in which agents aggregate their individual beliefs via majority voting. The model then checks whether the group opinions aggregated from individually consistent beliefs are also consistent. Inconsistencies are observed in different scenarios, but groups with high opinion diversity are particularly affected. This result can help us understand the impact of diversity in epistemic group decision-making, particularly in exceptional conditions involving uncertainty or pressure of time. I close by comparing this effect to previous results about diversity from other computational approaches.
Date: October 21, 2024
Speaker: David Freeborn
Title: Epistemic Factionalisation for Bayesian Agents with Multiple, Probabilistically Related Beliefs
Abstract: Epistemic polarisation occurs when agents’ beliefs move further apart despite receiving the same evidence, whilst epistemic factionalisation arises when groups form around multiple, correlated beliefs. I present a model of how a population of ideally Bayesian agents can divide into epistemic factions, even when all agents update on the same evidence This kind of factionalisation is driven by probabilistic relations between the agents' beliefs, with background beliefs shaping how the agents' beliefs evolve in the light of new evidence. Through this model, I examine conditions under which groups form stable factions and how these factions resist convergence despite access to shared information. This work offers insights into how and why beliefs about very different topics often coalesce within epistemic factions.
Date: November 4, 2024
Speaker: Sahar Heydari Fard
Title: Rethinking Diversity: The Role of Homogenous Clusters in Evolving Consensus
Abstract: Previous simulation models highlight the benefits of cognitive and demographic diversity for collectives. However, these models often assume a fixed amount of global diversity without considering its distribution, which often leads to the belief that maximal local intermixing is always beneficial. More importantly, these models usually assume minimal resistance to change, focusing on how consensus is reached rather than how it can evolve. This paper introduces a new model that challenges both these assumptions. It examines how the distribution of diversity affects group performance and explore scenarios where consensus can evolve despite significant resistance, such as when a minority group drives change against a majority consensus or in the presence of asymmetric conformity pressures or trust relations. The findings suggest that forming homogeneous clusters, as opposed to maximal intermixing, can be an effective strategy for minority groups to influence and improve the majority consensus.
Date: January 27, 2025
Speaker: Matthew Coates
Title: Multilayer Networks and the Evolution of Risky Cooperation
Abstract: Philosophers have shown that social network structure significantly influences the emergence of prosocial behaviors. Real social communities are often characterized by agents acting in different social spaces at the same time, such as family, friendship and work groups, distinguishing their behavior in each. However, network structures that capture this, such as multilayer networks, have received little study by philosophers researching the emergence of prosocial behaviors. In this work, I introduce multilayer networks to philosophy by demonstrating that they have a significant effect on the extent to which cooperation emerges in cases where cooperation is risky, using the stag hunt game. I show that multilayer structures increase the likelihood of cooperation emerging under certain conditions, and decrease it under others. Importantly, we can see the emergence of different strategies on different layers, despite those layers having no distinguishing features. Therefore, we can see the emergence of agents distinguishing their behavior in different social spaces even without strong assumptions about the meaning of those social spaces. Given many real social communities are multilayer, I consequently argue that multilayer networks should be investigated when studying the impact of networks on prosocial behaviors.
Date: February 10, 2025
Speaker: Nathan Gabriel
Title: Social identity signaling in the generalized Bach or Stravinsky game
Abstract: People have multifaceted social identities. In addition to this, people occupying the intersection of two identities can have preferences that are more than the sum of the preferences of those occupying just one of the two identities. This presentation shows how this can be captured in a formal model of a coordination game and shows simulation results. While the model does capture some aspects of intersectionality, it was not originally designed with this in mind. The model was initially designed to capture, in general contexts, the cultural evolution of social group structures. I am looking for feedback on how the model's connection with intersectionality can be clarified and strengthened.
Date: February 24, 2025
Speaker: Leon Assaad
Title: Modeling and Measuring Multi-Option Polarization
Abstract: Formal models from social epistemology aim to capture polarization as the outcome of deliberation among boundedly rational agents. Classical models focus on conversations about a central proposition being true or false (Hegselmann, Krause, et al. 2006; Olsson 2013) or which of two strategies is better (Zollman 2007; O’Connor and Weatherall 2018). Hence, agents typically face a choice between two options and may polarize accordingly. However, many real-world conversations take a different form: Which of n options is true? Let us call topics where n > 2 multi-option topics. This talk addresses two challenges: first, it presents a new model of rational deliberation on multi-option topics. Second, it investigates how increasing the number of discussed options affects potential polarization. The model extends Assaad and Hahn’s 2024 simulation of groups deliberating a topic via evidence exchange (a "NormAN" model, Assaad et al. 2023). It represents the topic as a variable with n mutually exclusive and jointly exhaustive “truth values.” Do more options (n) increase polarization? This depends on how we define multi-option polarization. Two measures are proposed: (1) the number of “positions” the population converges on, capturing the variety of favored options, and (2) belief dispersion, quantifying disagreement about each option. Preliminary results show that increasing n both exacerbates and reduces polarization: the number of defended positions rises, making it harder to settle on a single preferred candidate, while average belief dispersion decreases, reflecting less extreme disagreement about individual options. These results raise essential questions about the measurement of multi-option polarization.
Date: March 10, 2025
Speaker: Patricia Rich
Title: Trust Updating in Strongly Tied Networks
Abstract: We present a new agent-based modeling framework for studying how testimonial relationships, trust levels, and the reliability of primary information sources influence agents' beliefs. The model combines elements from network epistemology, sociology, and trust-updating models. We share the results of simulations run within this framework, and use them to address two questions: First, how should we characterize an agent's epistemic position within a network? Second, is polarization inevitable when both trust and belief evolve? In response to the first question, we use our results to support our favored characterization of epistemic position over an alternative in the literature; in brief, the reliability of an agent's trusted sources is a strong predictor of an agent's epistemic success, whereas having many diverse and independent sources is a very weak predictor. In response to the second question, we show that polarization may not loom as large as other modeling frameworks have suggested. (The talk is based on joint work with Sergiu Spatan)
Date: March 24, 2025
Speaker: Matteo Michelini
Title: Can Scientific Communities Profit from Evaluative Diversity?
Abstract: Current epistemic landscape models assume that agents agree on the epistemic value of every research approach. Yet, science is frequently characterized by a diversity of standards, where scientists may value the same approach differently depending on their specific research aims. Using computer-based simulations, we investigate how this diversity --which we term 'evaluative diversity'-- influences collective scientific inquiry in problem-solving scenarios. Our findings show that communities with diverse evaluative standards benefit significantly from a continuous exchange of information. For such communities to function effectively, it is essential for scientists to share all explored approaches, even those they consider as having low epistemic value based on their own standards. Moreover, our results suggest that a moderate degree of evaluative diversity enables scientists, under specific conditions, to achieve more rewarding outcomes than those seen in homogeneous communities. These insights bear important implications for philosophical discussions on scientific pluralism and the incentive system in science. (The talk would be based on joint work with Javier Osorio)
Date: April 7, 2025
Speaker: Marina Dubova
Title: Computational investigation of experimentation strategies in science
Abstract: Scientists must choose which among many experiments to perform. To investigate the effectiveness of experimental strategies proposed by philosophers of science and executed by scientists themselves, we developed a multi-agent model of the scientific process that includes active experimentation, theorizing, and social learning. Our findings suggest that randomly choosing new experiments leads the agents to develop the most accurate theories of their simulated environments. In contrast, agents aiming to confirm theories, falsify theories, or resolve theoretical disagreements by their experiments may develop promising accounts for the data they collected, but ultimately misrepresent the ground truth they intended to learn about.