The core finding of the research is that highly effective steering methods tend to achieve conditioning at the cost of fluency—the more powerfully a method manipulates model behavior, the more noticeably the naturalness and coherence of the generated text decline.