Dynamically Bifurcating Tree Networks: A Conceptual Architecture for Growth-and-Pruning–Based Sequence Models

Most neural sequence models — including the Transformer architectures behind current large language models — use a fixed computation graph: the number of layers, heads, and parameters is set before training and does not change based on what the model encounters. This paper sketches an alternative: a network built from simple recursive units, called neons, that can structurally grow (by splitting) or shrink (by pruning) during operation, so that the shape of the network itself is part of what it learns. A neon accumulates a bounded scalar state; when that state saturates, the unit splits into two child units that inherit the overflow; when a unit's state collapses below a floor, it and its descendants are removed and a record of the removal constrains future growth in that region of the structure. Units at similar structural positions interact multiplicatively, gating each other's state. We describe the mechanism formally, relate each piece to existing, established techniques in neural architecture search, memory-augmented networks, and positional encoding, and lay out the open problems that would need to be resolved before this could be implemented and tested. This is an idea paper: no experiments have been run, and the central claim is that the mechanism is worth prototyping, not that it works. This paper is one of five companion papers making up The Lovial Framework, applying that framework's general concept of relationship-first, recursively stabilizing structure to a concrete AI architecture proposal; here, the role played elsewhere by a Lovial is played by a neon in the process of splitting or being pruned.