Christian Hilbe

Sebastian Pohl, Harsh Mehta, Pranav Mambayil et al.

LLMs are increasingly deployed as proxies for human study participants in social science experiments, yet the fidelity of this practice has rarely been tested directly. We test whether six LLMs can simulate individual human belief updates, comparing LLM outputs 1-to-1 against ground truth data from 391 UK participants on Prolific, who updated their stances on three discussion topics after reading Reddit comments. Each participant was simulated by an LLM conditioned on a persona derived from their demographic and personality trait data. We find that some LLMs (Qwen3-32B and GPT-5-Mini) can match the human post-stance distribution, but only when given participants' actual initial stances. All six models fail to simulate initial stances themselves and to produce faithful belief updates from self-generated stances. Three systematic biases emerge across all models: overrepresentation of neutral positions, more frequent but smaller belief shifts than humans, and a failure to rank comments by convincingness. Demographic and personality trait personas had no consistent effect on fidelity. LLM simulations of human belief dynamics are only reliable when grounded in realistic starting conditions, that current multi-round social media simulations rarely provide.

Franziska Lesigang, Christian Hilbe, Nikoleta E. Glynatsi

In this paper, we study reactive strategies in repeated additive games between two players with finitely many actions. Reactive strategies condition only on the opponent's previous action, making them one of the simplest ways players can respond to past interactions. Additive games include important models of cooperation, such as the donation game and games with a punishment option. We show that, for this class of games and strategies, the conditions for symmetric Nash equilibria reduce to a system of linear equalities and inequalities in the strategy parameters, allowing us to characterise all such equilibria. We establish a one-to-one correspondence between non-empty subsets S of the action set and equilibrium classes, which we call S-supporting equilibria. These are equilibria that use exactly the actions in S when playing against themselves. As a special case, we recover the well-known equalizer strategies as the equilibria supported on the entire action set. To assess which equilibrium classes are most evolutionarily relevant, we complement our analytical characterisation with simulations of social learning dynamics. We find that their prevalence is determined by two factors: how likely they are to be generated and how robust they are against invasion.