Where Cognition Lives: Dissecting Emergent from Computed Function in a Minimal Complete Cognitive Architecture

This paper investigates the boundary between emergent cognitive functions—those that arise naturally from scale and gradient descent—and those that require explicit, engineered machinery. The authors build a minimal, complete cognitive architecture consisting of a recurrent reasoner, a homeostatic control field, and a value module to test whether specific cognitive functions (like adaptive halting and compute allocation) emerge or must be computed.

The authors employ a rigorous experimental protocol, including preregistration, adversarial review, and the use of positive controls for every null result. They test their architecture on algorithmic composition tasks where difficulty and information structure are precisely controlled. They also extend their analysis to a frozen 14B-parameter large language model (LLM) to evaluate test-time compute governance, such as self-consistency voting and inter-sample agreement.

The study reveals that while basic competence and adaptive halting emerge from training, the system fails to learn effective second-order decisions (i.e., deciding how much compute to allocate to a task). The authors find that gradient descent actively suppresses task-irrelevant information, making it impossible for the model to learn optimal allocation policies internally. Instead, these decisions must be computed by an explicit allocator using learned self-models. Furthermore, the authors demonstrate that the apparent advantage of "self-observation" in adaptive halting is largely an artifact of how mixture readouts are processed, rather than genuine cognitive insight. Finally, they show that in environments with "cliff-cost" structures—where the cost of computation is invisible until spent—explicit allocation provides significant value, confirming that the structure of the environment dictates the necessity of computed governance.