CALLIOPE · BCOM

Computational psychiatry is not a single modelling paradigm. It spans mechanistic dynamical models of the brain, cognitive process models, longitudinal symptom models, normative population models, and increasingly artificial-intelligence models trained to compress, generate, predict, or simulate high-dimensional neural and behavioural data. These approaches differ primarily in what they treat as the relevant state, which observations they fit, and which causal claims they support. We provide a compact map of this landscape, treating dynamic causal modelling as an inversion and model-comparison methodology within the broader family of whole-brain and neural dynamical models. We then position the Algorithmic Agent Model as a complementary architectural programme: the whole-brain model supplies the neural substrate, while the Agent Model specifies the world-modelling, valuation, planning, and action functions that a behaviourally competent brain model must implement. The resulting target is an agentic whole-brain model: a patient-specific, closed-loop model constrained jointly by anatomy, neural dynamics, behaviour, experience, and intervention response.

A blueprint for building brain models that actually behave like people, not just oscillate like brains.

Current whole-brain models (WBMs) are genuinely impressive: given a patient's connectome and resting neural recordings, they can reproduce the statistical fingerprint of brain activity — which regions co-activate, how fast oscillations are, how the system responds to a pulse of stimulation. But here's the problem the paper is picking at: a model that reproduces resting functional connectivity is a model of brain dynamics, not a model of a person. It doesn't have goals. It doesn't plan. It can't learn from consequences or develop the kind of maladaptive loops that define psychiatric illness.

The paper first maps the computational psychiatry landscape honestly. There are at least six distinct model families — mechanistic neural models, cognitive process models, symptom network models, normative deviation models, AI representation models, and large-scale foundation models — and they differ not just in method but in what they even treat as the relevant "state" of the system. Crucially, these aren't competing levels in a hierarchy; they're intersections of two independent axes: what you coarse-grain (neural populations, beliefs, symptoms, behavior) and how you build the model (dynamical simulation, Bayesian inversion, supervised learning, etc.). Dynamic causal modelling, for instance, is a methodology for inverting and comparing generative neural models — it's not a separate descriptive level. Getting this taxonomy right matters because it clarifies what each approach can and cannot support causally.

The paper's main architectural proposal is the Algorithmic Agent Model, which says a complete brain model needs four functional components: a Modeling Engine (building and updating a world model), an Objective Function (evaluating states and generating valence), a Planning Engine (searching over future trajectories), and an Action system. These aren't separate boxes wired to a brain model — they're proposed as identifiable collective coordinates of the same whole-brain dynamical system. The key move is embedding the WBM inside a closed agent-world loop, so the model's outputs change its future inputs. Without that loop, you can't reproduce learning, maladaptive policies, or the self-reinforcing cycles central to conditions like depression. For MDD specifically, the architecture distinguishes five mechanistically distinct failure routes: biased world modeling, abnormal valuation (anhedonia), failed planning, a genuinely adverse environment, and coupled self-reinforcing loops — each pointing toward different interventions.

The paper is also honest about what this demands methodologically. Resting data alone cannot discriminate between, say, pessimistic beliefs and reduced reward sensitivity producing the same withdrawal behavior. The programme requires designed perturbations — tasks targeting reward learning, controllability, effort discounting, affective forecasting — combined with simultaneous neural recordings and longitudinal outcomes. The validation target isn't just predicting brain signals under stimulation; it's predicting the joint distribution over brain activity, behavior, speech, and experience reports under interventions. A five-stage developmental ladder is sketched, from a basic resting WBM (Stage 0) up to a naturalistic communicative twin that predicts what a patient says and does, and how both change with treatment (Stage 4).