Since there is nothing special about the uniform (!), re-running the evaluation with a Jeffreys prior Dir(??,??,???,??) reduces this feature, which anyway is a characteristic of the prior distribution, not of the posterior distribution which accounts for the data.