Recent work by Wang (2025) argues that “AI = Dynamic Categorization,” framing language modeling as the real-time construction of temporary semantic prototypes that guide token selection. The Transformer implements this via self-referential attention, where Value (V) vectors—crucially—encode response semantics (“what should be said next”) rather than input semantics. However, this paradigm remains a one-step decision process, lacking the capacity for iterative refinement. In this paper, we propose DynaCat-Diffusion…