Evaluating Persona Consistency and Emotion Understanding in Small Language Models

As Large Language Models (LLMs) become increasingly widespread and integrated into everyday assistant applications, achieving a reliable behavior is critical for effective human–AI interaction. We argue that trustworthy AI assistants rely on two key dimensions: (i) persona consistency, i.e., maintaining a stable personality profile grounded in well-established trait descriptions, and (ii) emotion understanding, i.e., the ability to appropriately infer affective signals. In this work, we evaluate a set of Small Language Models (SLMs), a resource-constrained subclass of LLMs suitable for on-device deployment, along these two dimensions. For persona consistency, we assess how accurately models adhere to explicit persona constraints across varying prompt conditions. For emotion understanding, we evaluate how models recognize emotions and infer their causes across multiple evaluation settings, including conditions where an external emotion label is provided. Our experiments showed that persona consistency depends on prompt detail, while affective-signal injection improves emotion understanding, hence suggesting that trustworthy deployment of SLMs benefits from fine-grained persona instructions and an upstream emotion recognition module.