Authors: Jayd Matyas, William A. Cunningham, Alexander Sasha Vezhnevets, Dean Mobbs, Edgar A. Duéñez-Guzmán, Joel Z. Leibo
Organizations: Google DeepMind · University of Toronto · Caltech · Gibran AI
Abstract
Attitude change - the process by which individuals revise their evaluative stances - has been explained by a set of influential but competing verbal theories. These accounts often function as mechanism sketches: rich in conceptual detail, yet lacking the technical specifications and operational constraints required to run as executable systems. We present a generative actor-based modelling workflow for "rendering" these sketches as runnable actor - environment simulations using the Concordia simulation library. In Concordia, actors operate by predictive pattern completion: an operation on natural language strings that generates a suffix which describes the actor's intended action from a prefix containing memories of their past and observations of the present. We render the theories of cognitive dissonance (Festinger 1957), self-consistency (Aronson 1969), and self-perception (Bem 1972) as distinct decision logics that populate and process the prefix through theory-specific sequences of reasoning steps. We evaluate these implementations across classic psychological experiments. Our implementations generate behavioural patterns consistent with known results from the original empirical literature. However, we find that achieving stable reproduction requires resolving the inherent underdetermination of the verbal accounts and the conflicts between modern linguistic priors and historical experimental assumptions. We document how this manual process of iterative model "stabilisation" surfaces specific operational and socio-ecological dependencies that were largely undocumented in the original verbal accounts. Ultimately, we argue that the manual stabilisation process itself should be regarded as a core part of the methodology functioning to clarify situational and representational commitments needed to generate characteristic effects.
Language models can state that "the Earth orbits the Sun" and, when role-playing Aristotle, assert the opposite. Recent work argues that persona adoption is fundamental to how language models behave, with models selecting the most appropriate persona for a given context. Does such role-playing merely change the model's outputs, or does it also affect what the model internally represents as truthful? We study this question using the role-play of characters whose beliefs differ from the modern consensus, and induce personas with a number of different methods: prompting, in-context learning (ICL), supervised fine-tuning (SFT), and Open Character Training (OCT), and Emergent Misalignment (EM). We measure belief internalization across these approaches with truth probes and with behavioral tests, finding a broad spectrum of belief internalization. Prompting, ICL, and SFT change what the model says with little representational change. EM creates a large, broad shift in the model's truth representation, and OCT a smaller shift that is clearest on the larger model. Understanding when training changes a model's worldview rather than merely its behavior may become increasingly important as AI systems are entrusted with greater autonomy and influence.
People make decisions differently in strategic interactions. Some update beliefs like a Bayesian; others exhibit biases like motivated reasoning. Although creators of large language models use simulated humans for safety evaluations and training, they often fail to cover this breadth of human behavior. We argue that cognitive science and economics provide a convenient tool for doing so, making use of mathematical models of human decision-making. We propose an approach that we call Equation-to-Behavior Prompting for guiding large language models to match cognitive models, and evaluate this approach on persuasion games based on legal decision-making. We find that large models can approximate equation-based specifications -- Bayesian updating, affine distortion, motivated updating, and Grether's α-β model -- using prompting, but small models fail to do so. However, training small models with reinforcement learning to adhere to mathematical rules, Equation-to-Behavior RL, reduces belief error by 26.5% in out-of-distribution parameterizations. We show that these simulations can help create diverse training environments; training small models to consider different kinds of decision-makers improves average belief change by 2.5%--12% over Bayesian-only training, even when persuading GPT-5-mini. Our work could improve human simulations for training and evaluation in increasingly realistic settings, and could also enable novel research into more complicated mathematical models of human decision-making.
Activation steering promises modular control of language model behavior: a behavior such as politeness corresponds to a direction in a model's activations, and adding that direction while it generates should switch the behavior on and leave everything else alone. It does not. We ask what decides which other behaviors move, and by how much, and find that it is the model rather than the behavior being steered. A steer relaxes the model toward a small set of behaviors it already favors, chiefly refusal, sycophancy, and poeticism, and that set is much the same whatever is steered. Three results across 24 behaviors and ten instruction-tuned models support this, every effect read off the generated text by a language-model judge rather than off a probe. That readout matters: all 24 behaviors are linearly decodable, but only 20 change what the model writes. First, a direction carrying no behavioral content, matched to a real steer only in the size of the vector it adds, moves the same behaviors in the same order as real steers do, while producing none of the behaviors that need a specific direction. Second, most interference runs one way, so it cannot be an overlap between two directions: steering profanity makes the model toxic, while steering toxicity leaves profanity untouched. Third, with a behavior held out entirely, geometry measured on the others explains almost none of the interference it takes part in. The account holds on all ten models, the pull toward defaults strongest below 10B parameters and weakening in each family's largest. Reading a steer as a perturbation whose endpoint the model fixes implies that disentangling behavior directions cannot by itself make steering modular.