In the architecture proposed by Wiesław L. Galus and Janusz A. Starzyk, a hungry system does not just calculate the highest reward. It weighs need thresholds, shifting goals, the state of its body, its remaining resources, and its uncertainty about whether an action will even work. The authors argue this makes their approach a more adequate account of how an embodied system chooses a response under regulatory pressure than standard reinforcement learning, which optimizes reward without those constraints.
The stakes here are conceptual rather than empirical. Most machine-learning agents treat the body as an afterthought and reward as the sole currency. This paper asks what changes when interoception, the sense of one's own internal state, is written into the mathematics from the start. If motivation and affect are part of the substrate rather than decoration on top of it, then the same formal machinery that governs a robot's hunger might bear on how any embodied system generates awareness of itself.
The model's central move is a re-entrant loop. The authors formalize how feedforward processing, lateral interactions between units, and feedback pathways combine, along with the selection mechanisms that decide which representations win. Incoming exteroceptive and interoceptive signals, bodily-motivational context, and memory traces are bound into associative structures the authors call semblions. These semblions compete for access to further processing and for top-down reconstruction. Above this competition sits what the authors describe as global affect, a single control signal that tunes the learning rate, the valence assigned to representations, and the balance between exploration and exploitation.
What the paper does not do is test any of this. The method is mathematical and theoretical formalization. The abstract reports no experiments, no simulations, and no empirical data. The claim that motivated learning outperforms standard reinforcement learning is made by assertion, not by measurement. The definitions of semblions and the role of affect are stated as modeling choices, and the authors present the work as a step toward rigor and a possible basis for later simulation and biologically inspired AI implementation. It is also a preprint, not yet peer reviewed.
Still, the framing sharpens a question worth holding open. The authors position affect and interoception as the substrate from which adaptive awareness and self-regulation emerge. If consciousness is bound up with a system representing and regulating its own bodily state, then a formal loop that binds interoceptive signals into competing representations is either a map toward that experience or a demonstration of how far equations can go without touching it. This model does not tell us which. It only makes the boundary easier to see.
Source: arXiv:2609.20437, http://arxiv.org/abs/2609.20437v1