Personalization of large language models is increasingly important, but most existing approaches adapt dense LLMs. According to a new arXiv preprint, these dense models are hard to steer because their neurons are polysemantic—each one encodes multiple unrelated concepts, which muddies the effect of fine-tuning or prompting.

The paper instead focuses on mixture-of-experts (MoE) LLMs, where different experts handle different functions. The authors aim to explain the roles of these experts in the context of personalization, suggesting that the modular structure may make it easier to identify which components should change for a given user.

Because the abstract is truncated, the full method and results are not yet visible. The contribution, as described, is an explanatory framework for expert roles rather than a new personalization algorithm per se.