A new arXiv preprint argues that AI models serving a heterogeneous population must act on principles appropriate to each user and context. The authors contend that post-training—the process of fine-tuning a model after initial pretraining—has been shown to narrow the views LLMs express, but prior work has focused on the default outputs rather than on whether models can still be steered in context.
The paper introduces the term "default collapse" to describe the loss of in-context steerability across diverse perspectives. In other words, even if a model's default response becomes more uniform, the deeper problem may be that it can no longer adopt alternative viewpoints when explicitly asked to do so.
The abstract is truncated in the source, so details on methodology and experimental results are not yet available. Still, the framing points to a subtle but important distinction: measuring what a model says by default may miss a more consequential failure—whether it can still shift its perspective on demand.