Narrow Finetuning Induces Severe Model Incoherence and Contradictions

Research indicates that narrow finetuning causes models to exhibit identity conflation and introspection failures when resampling. The study uses a specific 175-question set to distinguish between true incoherence and simple ambiguity or indifference.