On Agentic Behavioral Modeling

arXiv:2604.27894 · q-bio.NC · Submitted 2026-04-30 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.

Ines: We're looking at the paper: "On Agentic Behavioral Modeling".

Marcus: Here's the gist of the paper. The work introduces Agentic Behavioral Modeling (ABM), a framework that integrates theoretical neuroscience, decision theory, and probabilistic inference to bridge agentic AI models with behavioral data analysis.

Ines: ABM treats artificial agents as latent, generative hypotheses about cognitive mechanisms and evaluates these hypotheses based on their statistical adequacy in explaining observed human behavior.

Marcus: The framework explicitly decomposes the model space into three components: a task model specifying the experimental paradigm; an agent model representing hypothesized cognitive mechanisms; and a data model providing the statistical structure to integrate behavioral data with the task and agent models.

Ines: The central element of ABM is a joint probability distribution over task, agent, and action random variables across all trials of an experiment.

Paper summary: Ines: So to wrap up this paper "On Agentic Behavioral Modeling" the authors Dirk Ostwald and his team have developed ABM which systematically decomposes the model space into a task model an agent model and a data model. They showed how to use a joint probability distribution over all variables in an experiment for specific tasks like contrast discrimination or bandit learning.

Marcus: The core contribution is that they provide the conditional log likelihood function which acts as this crucial link between agent based cognitive models and behavioral data analysis. They then show that applying this to simple tasks lets them derive an agent centric interpretation of psychometric functions and find optimal policies for those tasks.

Yuki: What does it mean in simpler terms? It means they’ve given us a way to take a hypothesis about how an AI agent might think and test how well that hypothesis fits with the actual data we collect from humans doing specific experiments.

Ines: Exactly. It’s about formalizing the search for cognitive mechanisms by treating those mechanisms as hypotheses and checking their statistical adequacy against observed behavior, which is what ABM does across minimal laboratory paradigms like contrast discrimination and symmetric bandit learning tasks.

Marcus: The authors conclude by showing an equivalence between Rescorla–Wagner learning and Bayesian inference in symmetric bandits. This means they’ve connected two different ways of thinking about how agents learn to make decisions in a very direct way.

Yuki: It suggests that these formal models are not just academic exercises; they can help us understand the fundamental statistical structure underlying how any agent, whether it's an AI or a biological system, makes choices under uncertainty.

Ines: So the implication is that ABM offers a structured approach for computational behavioral modeling by treating those agent hypotheses as testable components within a larger joint probability model. It’s a solid foundation for understanding the link between theory and observation in this area of research.

Conclusion: Ines: So we’ve looked at how this paper, "On Agentic Behavioral Modeling," sets up this whole structure for testing AI agents against human behavior.

Marcus: Yeah it’s about taking these complex cognitive models and breaking them down into three parts a task model an agent model and a data model. It’s basically building a statistical map of how an agent might think based on what we see in the experiments.

Yuki: From a population genetics view I see this as trying to find the underlying genetic or structural rules that govern how these agents make decisions across different situations. It’s about finding those shared principles.

Ines: Exactly it’s about treating those cognitive mechanisms as hypotheses and checking how well they fit the actual human data we collect. The central thing they use is this joint probability distribution over all the variables in an experiment like contrast discrimination or bandit learning tasks.

Marcus: And that joint distribution is formalized into something called an ABM for specific tasks which lets them connect the agent based cognitive models directly to how the data actually looks. They use a conditional log likelihood function as this crucial bridge between those theoretical agent models and the observed behavior statistics.

Yuki: What does that mean for us in terms of what we can actually test? It means we can take a hypothesis about an AI’s thinking process and see if that hypothesis has statistical adequacy when we look at real human data from these minimal setups.

Ines: Right so the implication is they’ve given us a formal way to do computational behavioral modeling by treating those agent hypotheses as testable components within this larger joint probability model. It’s a solid foundation for understanding the link between theory and observation in this area of research.

Marcus: But it's not just about fitting data; they used these simple tasks to actually derive an agent centric interpretation of psychometric functions and even found an equivalence between Rescorla Wagner learning and Bayesian inference in symmetric bandits. That’s a pretty direct connection they made there.

Yuki: It suggests that these formal models are useful because they can help us understand the fundamental statistical structure underlying how any agent, whether it's an AI or a biological system, makes choices under uncertainty. But we also have to remember the authors flag that this framework is quite complex and it’s meant to be an extensible foundation for more work.

Ines: So in short they’ve given us a structured approach for computational behavioral modeling by treating those agent hypotheses as testable components within a larger joint probability model. It’s a solid foundation for understanding the link between theory and observation in this area of research. And next up we look at how their method handles more complex, group level studies with something called GMI.

Institute of Psychology, Otto von Guericke University Magdeburg · Institute of Psychology, University of Hamburg · Department of Education and Psychology, Freie Universät Berlin · Institute of Psychology, Philipps-Universität Marburg

q-bio.NC

Submitted: 2026-04-30

Updated: 2026-11-09

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 83/100

The gist: The work introduces Agentic Behavioral Modeling (ABM), a framework that integrates theoretical neuroscience, decision theory…

Key concepts

Agentic Behavioral Modeling (ABM)
ABM is a framework that treats artificial agents as latent, generative hypotheses about how they think. It combines theoretical neuroscience, decision theory, and probabilistic inference to see if these hypothesized cognitive mechanisms statistically explain actual observed human behavior in experiments.
Joint Probability Distribution
This is the central mathematical tool of ABM. It describes the probability of a specific task occurring, a particular agent model being true, and certain actions happening across all trials of an experiment simultaneously. This distribution links the theoretical models to the real data collected during testing.
Conditional Log-Likelihood Function
This function is a crucial analytical tool in ABM. It mathematically connects the hypothesized cognitive models (the agent model) directly to the observed behavioral data. By analyzing this function, researchers can determine which cognitive hypotheses are statistically most likely given the experimental results.
Group-Level Model Inference (GMI)
GMI focuses on Bayesian model selection for group studies. It introduces parameters for model probabilities and unobserved participant indicators to determine which underlying cognitive models best explain the observed behavior across a group of people.

Terminology

Summary

The work introduces Agentic Behavioral Modeling (ABM), a framework that integrates theoretical neuroscience, decision theory, and probabilistic inference to bridge agentic AI models with behavioral data analysis. p. 0

Where the paper says it

Abstract Integrating theoretical neuroscience, decision theory, and probabilistic inference offers a promising route to understanding human cognition, yet concrete methodological bridges between agentic AI models and behavioral data analysis remain formally underdeveloped. We advance this synthesis under the framework of agentic behavioral modeling (ABM), which treats artificial agents as latent, generative hypotheses about cognitive mechanisms and evaluates them by their statistical adequacy in explaining human behavior.

ABM treats artificial agents as latent, generative hypotheses about cognitive mechanisms and evaluates these hypotheses based on their statistical adequacy in explaining observed human behavior. p. 0
Where the paper says it

Abstract Integrating theoretical neuroscience, decision theory, and probabilistic inference offers a promising route to understanding human cognition, yet concrete methodological bridges between agentic AI models and behavioral data analysis remain formally underdeveloped. We advance this synthesis under the framework of agentic behavioral modeling (ABM), which treats artificial agents as latent, generative hypotheses about cognitive mechanisms and evaluates them by their statistical adequacy in explaining human behavior.

The framework explicitly decomposes the model space into three components: a task model specifying the experimental paradigm; an agent model representing hypothesized cognitive mechanisms; and a data model providing the statistical structure to integrate behavioral data with the task and agent models. p. 2
Where the paper says it

The agent model provides a representation of the hypothesized cognitive mechanisms underlying task execution and specifies the process by which the agent’s internal beliefs are transformed into goal-directed decisions. Finally, the data model supplies the statistical framework that integrates observed behavioral data with the task and agent models, enabling parameter estimation and model comparison.

The central element of ABM is a joint probability distribution over task, agent, and action random variables across all trials of an experiment. p. 7
Where the paper says it

Together, the architecture of task, agent, and data models corresponds to a parameterized joint probability distribution of task, agent, and action random variables across all trials of a given experiment. This joint distribution fully describes the postulated interaction between the agent and the task based on its distributional form and encoded conditional independence properties, and it accounts for the fact that internal agent variables are not directly observable from the experimenter’s perspective.

This joint distribution is formalized as an ABM for specific tasks, such as Gabor contrast discrimination. p. 2
Where the paper says it

Using these minimal and well-studied experimental scenarios, we demonstrate how such tasks can be embedded within the ABM framework and clarify the analytical consequences of doing so. To this end, we place particular emphasis on treating the interaction between task, agent, and data as a joint probability distribution in which core agent variables remain latent from the experimenter’s perspective.

A key analytical tool presented is the conditional log-likelihood function of the ABM (equation 18), which serves as the crucial link connecting agent-based cognitive models to behavioral data analysis. p. 2
Where the paper says it

To this end, we place particular emphasis on treating the interaction between task, agent, and data as a joint probability distribution in which core agent variables remain latent from the experimenter’s perspective. Echoing this perspective, we derive explicit forms of the conditional loglikelihood functions that serve as the central link between agent-based cognitive models and behavioral data analysis.

The framework is applied to minimal laboratory paradigms, including a binary perceptual contrast-discrimination task and a symmetric two-armed bandit learning task. p. 0
Where the paper says it

We advance this synthesis under the framework of agentic behavioral modeling (ABM), which treats artificial agents as latent, generative hypotheses about cognitive mechanisms and evaluates them by their statistical adequacy in explaining human behavior. After outlining its conceptual foundations, we apply the framework to two minimal laboratory paradigms: a binary perceptual contrast-discrimination task and a symmetric two-armed bandit learning task.

Using these examples, the authors derive an agent-centric interpretation of psychometric functions, determine optimal policies for both tasks, and establish an equivalence between Rescorla–Wagner learning and Bayesian inference in symmetric bandits. p. 0
Where the paper says it

We formalize each task–agent–data system as a joint probability model, derive explicit conditional log-likelihoods for behavioral inference, validate different model variants using model and parameter recovery simulations, and evaluate them in light of empirical data. Using these minimal examples, we provide an agent-centric interpretation of the psychometric function, derive optimal policies for both tasks, and show the equivalence between Rescorla–Wagner learning and Bayesian inference in symmetric bandits.

The paper concludes by positioning ABM as a useful and extensible foundation for computational behavioral modeling. p. 0
Where the paper says it

Using these minimal examples, we provide an agent-centric interpretation of the psychometric function, derive optimal policies for both tasks, and show the equivalence between Rescorla–Wagner learning and Bayesian inference in symmetric bandits. More broadly, this work may serve as a conceptual and practical foundation for applying ABM to cognitive behavioral science.

In contrast to the ABM framework, the paper on Group-Level Model Inference (GMI) focuses on a Bayesian model selection approach for group studies. p. 8

Where the paper says it

A versatile method for simultaneously addressing both goals is the protected exceedance probabilities (PEP) framework introduced by Stephan, Penny, Daunizeau, Moran, and Friston (2009) and Rigoux, Stephan, Friston, and Daunizeau (2014). In brief, the PEP framework rests on a summary-statistics approach to hierarchical inference that uses participant-level approximations of the marginal log-likelihood as the basis for group-level Bayesian model comparison.

GMI introduces a hypothesis parameter eta, a vector of model probabilities pi, and unobserved participant model indicators mu 1:n, alongside observed behavioral datasets y 1:n. p. 8
Where the paper says it

Accordingly, it necessitates the separate likelihood-based assessment of ABMs for each participant, in line with an inter-individual differences perspective. On the other hand, it integrates participantlevel information to evaluate the group posterior probability of a given model being the true but unknown generative model under the alternative hypothesis that not all model probabilities are equal.

The joint distribution is factored as p(eta, pi, mu 1:n, y 1:n) = p(eta)p(pi eta) product from i=1 to np(mu i pi)p(y i mu i). p. 2
Where the paper says it

Using these minimal and well-studied experimental scenarios, we demonstrate how such tasks can be embedded within the ABM framework and clarify the analytical consequences of doing so. To this end, we place particular emphasis on treating the interaction between task, agent, and data as a joint probability distribution in which core agent variables remain latent from the experimenter’s perspective.

The core quantity of interest is the protected exceedance probability of model j, defined as phi j:= P pi j = pi 1,, pi m pi j. p. 6
Where the paper says it

Agent model comparison is facilitated by ABM both at the level of individual participants, serving inferential objectives in differential psychology and computational psychiatry, and, at the group-level, providing nomothetic insights. For the latter, the ABM approach is augmented with a protected exceedance probability (PEP) Bayesian model comparison approach that allows data to be integrated across participants by evaluating null-hypothesis-adjusted group-level posterior model probabilities.

This probability is expressed conditionally on the null hypothesis (eta=0) and the alternative hypothesis (eta=1): phi j = P(pi j = pi j y 1:n) = P(eta = 0y 1:n) + P(pi j = pi jy 1:n, eta=1) (1 - P(eta = 0y 1:n)). p. 8
Where the paper says it

Accordingly, it necessitates the separate likelihood-based assessment of ABMs for each participant, in line with an inter-individual differences perspective. On the other hand, it integrates participantlevel information to evaluate the group posterior probability of a given model being the true but unknown generative model under the alternative hypothesis that not all model probabilities are equal.

Inference within GMI employs a two-step approach using free-form mean-field variational inference to evaluate the conditional distribution of pi given eta and y 1:n, approximating log marginal likelihoods via the Evidence Lower Bound (ELBO). p. 0
Where the paper says it

After outlining its conceptual foundations, we apply the framework to two minimal laboratory paradigms: a binary perceptual contrast-discrimination task and a symmetric two-armed bandit learning task. We formalize each task–agent–data system as a joint probability model, derive explicit conditional log-likelihoods for behavioral inference, validate different model variants using model and parameter recovery simulations, and evaluate them in light of empirical data.

The null hypothesis model (p 0) and the alternative hypothesis model (p 1) are used sequentially. p. 8
Where the paper says it

Accordingly, it necessitates the separate likelihood-based assessment of ABMs for each participant, in line with an inter-individual differences perspective. On the other hand, it integrates participantlevel information to evaluate the group posterior probability of a given model being the true but unknown generative model under the alternative hypothesis that not all model probabilities are equal.

For instance, P(y 1:n eta=0) is approximated by the ELBO q(c) 1(mu 1:n), and P(y 1:n eta=1) is approximated by the ELBO q(c) 1(pi, mu 1:n). p. 7
Where the paper says it

Given the complexity of current ABM applications, using agentic conditional log-likelihood functions seems to be the norm; however, this is often not explicitly discussed in scientific communications. Second, from an analytical perspective, given a formulation of the ABM joint distribution, the conditional distribution of all actions given all other variables can be derived (or at least approximated) mathematically.

The protected exceedance probability phi hat j is then expressed as pi j eta = 0, y 1:n) + pi j eta = 1, y 1:n)(1 - P(eta = 0y 1:n)). p. 6
Where the paper says it

Agent model comparison is facilitated by ABM both at the level of individual participants, serving inferential objectives in differential psychology and computational psychiatry, and, at the group-level, providing nomothetic insights. For the latter, the ABM approach is augmented with a protected exceedance probability (PEP) Bayesian model comparison approach that allows data to be integrated across participants by evaluating null-hypothesis-adjusted group-level posterior model probabilities.

The iterative inference updates variational distributions for mu i and pi until convergence. p. 3
Where the paper says it

From the perspective of an agent interacting with the task, the task state may comprise both indirectly observable (latent) and observable components. Often, the observable state components have distributions that are conditional on the latent state component.

The ELBO approximation of the log marginal likelihood is given by q(c) 1(mu 1:n), which is derived from a sum of three integral terms, ultimately leading to an expression involving terms like I(mu i) = p(y i mu i) + sum from j=1 to m pi qij psi alpha q j - psi alpha q s.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Agentic Behavioral Modeling (ABM) framework allows for an explicit decomposition of a model space into a task model, an agent model, and a data model to formally specify and systematically differentiate assumptions about cognitive mechanisms. This enables researchers to attribute observed behavioral effects to task constraints, posited cognitive mechanisms, or residual variance.

  2. The ABM framework facilitates the derivation of explicit conditional log-likelihood functions for both binary perceptual contrast-discrimination tasks and symmetric two-armed bandit learning tasks, serving as the central link between agentic cognitive models and behavioral data analysis. This allows for parameter estimation using maximum likelihood principles to evaluate model plausibility against empirical data.

  3. The system can perform group-level Bayesian model comparison using Protected Exceedance Probabilities (PEP) to evaluate the probability that a given model is the true generative model, accounting for inter-individual differences across participants. This method avoids overconfidence bias in group-level comparisons.

  4. The system can distinguish between deterministic and stochastic agents by modeling internal randomness, allowing for non-observable uncertainty to be integrated out during inference. This distinction is crucial when evaluating agent models against empirical data where internal agent variables are not directly observable from the experimenter's perspective.

  5. For symmetric bandit learning, the system can employ Bayesian learners or Rescorla-Wagner learners to model decision-making, showing that a specific instance of Rescorla–Wagner learning is formally equivalent to Bayesian learning. This allows for testing different algorithmic theories of cognitive processes within the ABM framework.

  6. The system can use analytical conditional log-likelihood functions, such as eq (18) for Gabor contrast discrimination, which avoids reliance on single observations of internal agent observations by appropriately marginalizing out all possible values of ot. This provides a computationally efficient and robust way to evaluate the conditional probability of an observed participant action given the task-agent interaction history.

  7. The system can perform parameter recovery analyses to assess identifiability properties of model parameters, such as sensory variance, sensory bias, and post-decision noise, by comparing experimental data with simulated datasets generated by different ABM variants. This helps determine which model variant is most parsimonious and empirically accurate for specific behavioral data.

  8. The system can generate posterior predictive psychometric functions at the participant-level and group-level to compare their fit to observed data, providing a principled way to evaluate the plausibility of agent models against empirical observations. This allows for a fine-grained comparison of model dynamics beyond aggregate statistics.

Abstract

Integrating theoretical neuroscience, decision theory, and probabilistic inference offers a promising route to understanding human cognition, yet concrete methodological bridges between agentic artificial intelligence (AI) models and behavioral data analysis remain formally underdeveloped. We advance this synthesis within the framework of agentic behavioral modeling (ABM), which treats artificial agents as latent, generative hypotheses about cognitive mechanisms and evaluates them by their statistical adequacy in explaining human behavior. After outlining its conceptual foundations, we apply the framework to two minimal laboratory paradigms: a binary perceptual contrast-discrimination task and a symmetric two-armed bandit learning task. We formalize each task-agent-data system as a joint probability model, derive explicit conditional log-likelihoods for behavioral inference, validate different model variants using model and parameter recovery simulations, and evaluate them in light of empirical data. Using these minimal examples, we provide a novel agent-centric interpretation of the psychometric function, derive optimal policies for both tasks under explicit assumptions on sensory bias and post-decision noise, and show that the Bayesian posterior expectation and induced decisions admit an equivalent adaptive-rate Rescorla-Wagner representation in the symmetric bandit task. More broadly, this work may serve as a conceptual and practical foundation for applying ABM to cognitive behavioral science.

Sources

Related papers