We fantastic-tune virtual DMs with agent-generated and real interactions to evaluate expressiveness, and gauge informativeness by comp