These experiments start from the released NCM generator after its original two training stages. The question is whether further training or a changed generation schedule improves the generated text under comparable conditions.
Approach
The work reuses the upstream transformer, tokenizer, quantizer, and reconstruction operators. It compares a continuation of the released baseline objective with variants involving generator/decoder training, within-rank generation, and repeated refinement.
The current continuation study keeps the tokenizer and decoder frozen. A related refinement study changes how the generator revisits its intermediate codes. Longer runs preserve the optimizer, sampling state, and objective so that duration can be examined without silently changing the experiment.
The project does not reproduce the first two training stages from scratch. Comparisons use one training seed, and equal data exposure does not imply equal compute when one method performs additional passes.
Text quality and the effect of refinement need direct evaluation. Improvement on a code prediction or reconstruction objective alone would not settle either question.