It's not that different from how LLM tokens work, only in a tree structure as opposed to a plain sequence. Having a tree structure makes it easier to formally define rewrite rules (which is key for interpretability), as opposed to learning them from data as LLM do.