Language models typically learn content and style together, so the stylistic variation in their outputs is difficult to identify or steer. A new arXiv preprint asks whether these recurring styles can be discovered without any supervision, and if so, whether they can then be controlled.
The abstract describes a study of unsupervised style discovery, though the full method is not detailed in the available text. The authors appear to be probing whether stylistic patterns emerge naturally from model behavior, rather than requiring explicit labels or hand-crafted features.
If successful, such an approach would let users separate what a model says from how it says it, opening the door to more precise control over tone, formality, or other stylistic dimensions. The work is preliminary, but it points toward a more flexible way to manage the expressive range of language models.