What do you mean by iterative chunking? There are Transformer variants that hierarchically build token- or word-level representations from bytes, then reverse it to decode, but they never really took off. Didn't help much compared to a good tokenizer.