VocabX separates cheaper output from compressed decisions
Dharma-AI's essay-grading example makes vocabulary adaptation a question of decision quality as well as token count.
A smaller token count does not, by itself, establish a cheaper inference workload. Dharma-AI's VocabX article, published October 21, 2026, selects vocabulary elements using the different costs of processing input and generating output, rather than frequency alone.
Its main example is a Portuguese essay grader built on Qwen3-4B. The method rebuilds subword units and assigns dedicated tokens to recurring spans, including JSON structures. The comparison holds the base model, training data and training recipe constant.
One unsuccessful variant matters more than the headline savings. Combining a grade with its surrounding structure reduced evaluation quality. Keeping the grade separate produced a better result. Predictable formatting and the decision being assessed should therefore be treated as distinct compression targets.
For an independent evaluation, I would track structural validity, grading quality and measured serving cost separately. The authors say the underlying paper remains under review. The unresolved question is how well the advantage survives changes to the response schema and the essays being graded.
Source: Bernardo França de Aguiar and coauthors at Dharma-AI, in their account of vocabulary adaptation with VocabX.