VocabX cuts Qwen3-4B grader inference cost 77.0%

The starting point for VocabX was a production grader built on Qwen3-4B, a 4-billion-parameter model already trained for the task but still carrying the original Qwen3-4B tokenizer, a vocabulary of 151,669 entries built for many languages and kinds of text. Used for one language and one response format, that vocabulary split the JSON envelope and competency names into many pieces while scoring thousands of entries the task almost never uses. The method's design follows from how inference works: generating a token is memory-bound and costlier than reading one, and in the grader a generated token cost 7.4 times a token read. VocabX therefore rebuilds both how domain words are cut into pieces and which whole spans become single tokens, under a cost criterion rather than a frequency one. It also estimates the gain of candidates absent from the sampling sample, which raised token reduction on a contracts corpus from 44.7% to 52.8%. In the grader, an early vocabulary that fused each competency's grade with the surrounding JSON into one macro token was the most compact option but scored 0.689, below the production grader's 0.703; giving the grade its own token raised the bench score to 0.810 and cut cost by 77.0%, against 73.3% for the fused version. Across five public domains and three model families, VocabX compressed more than the strongest reference method in all fifteen combinations.