MotivatAI
MotivatAI
Explainable Motivational Control for Agents
A psychologically grounded, control-theoretic core engine for predictable, adaptive and auditable agent behavior.
Prof. Dr. Markus Quirin & Team
A White-Box Core Engine for Agent Behavior
Current AI agents can generate remarkably fluent responses, but their motivational and interpersonal behavior is largely emergent. Personality can drift, reactions can change unpredictably, and it is often difficult to explain why an agent behaved in a particular way.
MotivatAI takes a different approach.
At its core is an explicit computational model of psychological regulation. Needs, appraisals, incentives, perceived control, reward and threat, approach and avoidance tendencies, and their dynamics over time are represented as interacting state variables and feedback processes.
The aim is not to replace large language models. The core engine determines why and how an agent should act; generative AI can then support perception, language generation and interaction.
This separation makes behavioral dynamics interpretable, parameterizable, testable and, ultimately, auditable.
From Psychological State to Behavior
State Dynamics
The engine maintains explicit internal states for the agent and, where appropriate, estimates relevant states of the interacting user.
Control and Regulation
Psychologically grounded control processes continuously update needs, motivational tendencies, appraisals and behavioral priorities in response to internal and external events.
Behavioral Policy
The resulting state determines higher-level behavioral tendencies such as approach, withdrawal, support, challenge, exploration or boundary setting.
AI Interaction Layer
These behavioral specifications can be passed to an LLM or another agent architecture for linguistic or embodied execution.
The result is a hybrid architecture in which generative AI provides flexibility while an explicit psychological model provides behavioral structure and causal transparency.
Building the Causal Machinery Behind Behavior
Human motivation is not generated by a single variable. Multiple regulatory systems interact over time, often nonlinearly.
A mechanism that produces plausible behavior in isolation may create oscillation, escalation, deadlock or unrealistic dominance when coupled with other mechanisms. Different parameter combinations may also produce similar outcomes, while small changes can sometimes generate qualitatively different behavioral trajectories.
Our research therefore follows an iterative cycle:
psychological theory → formalization → computational implementation → simulation → behavioral testing → model revision
The central research challenge is to determine whether psychologically plausible regulatory mechanisms can be integrated into a stable, adaptive and interpretable computational architecture without delegating the underlying motivational dynamics to a black-box model.
Applications
The architecture is designed to remain independent of any single application domain or foundation model.
AI Companions
Persistent and adaptive interpersonal behavior over extended interactions.
Coaching/ Tutoring
State-sensitive support, challenge and behavioral adaptation.
Customer Interaction
Predictable behavioral policies, de-escalation and consistent interaction styles.