From a young age, we can select actions to achieve desired goals, infer the goals of other agents, and learn causalrelations in our environment through social interactions. Crucially, these abilities are productive or generative: for instance,we can impute desires to others that we have never held ourselves. This capacity has been captured by the powerful BayesianTheory of Mind formalism, but it remains to forge connections to the rich neural data around action selection, goal inference,and social causal learning. How can productive inference about actions and intentions arise within the neural circuitry of thebrain? Using the recently-developed linearly solvable Markov decision process, we present a neural network model whichpermits a distributed representation of tasks. Such a representation allows the expression of infinite possibilities by combininga finite set of bases, enabling truly generative inference of actions, goals, and causal relations in a neural network framework.

From a young age, we can select actions to achieve desired goals, infer the goals of other agents, and learn causal relations in our environment through social interactions. Crucially, these abilities are productive and generative: we can impute desires to others that we have never held ourselves. These abilities are often captured by only partially overlapping models, each requiring substantial changes to fit combinations of abilities. Here, in an attempt to unify previous models, we present a neural network underpinned by the linearly solvable Markov Decision Process (LMDP) framework which permits a distributed representation of tasks. The network contains two pathways: one captures the desirability of states, and another encodes the passive dynamics of state transitions in the absence of control. Interactions between pathways are bound by a principle of rational action, enabling generative inference of actions, goals, and causal relations supported by gradient updates to parts of the network.

One of the most salient and well-recognized features of humangoal-directed behavior is our limited ability to conduct mul-tiple demanding tasks at once. Previous work has identifiedoverlap between task processing pathways as a limiting fac-tor for multitasking performance in neural architectures. Thisraises an important question: insofar as shared representationbetween tasks introduces the risk of cross-talk and thereby lim-itations in multitasking, why would the brain prefer shared taskrepresentations over separate representations across tasks? Weseek to answer this question by introducing formal considera-tions and neural network simulations in which we contrast themultitasking limitations that shared task representations incurwith their benefits for task learning. Our results suggest thatneural network architectures face a fundamental tradeoff be-tween learning efficiency and multitasking performance in en-vironments with shared structure between tasks.