Neither Over- nor Under-Constrained: Evaluating LLM Personalization under a Constraint Qualification Audit.

arXiv:2610.25676v1 Announce Type: new Abstract: Accurate evaluation of LLM personalization is hindered by current evaluation paradigms that over- or under-constrain the response space, yielding frequent false rejections and false acceptances. To reduce these f