Simulating Multipersona Cultural Interaction: LLM Personas for AI Alignment

With the rise of large-scale synthetic persona data based in population census and demographics, it has become feasible to embody representative demographic and cultural contexts in LLMs as "personas". Recent research has demonstrated positive results on personas in reasoning capabilities, and rapid growth in open-source persona data and cultural benchmarking represent rich potential for improving AI cultural common-sense alignment and sovereign AI systems. We propose a "multipersona" prompting and multistep interaction framework for culturally grounded personas from NVIDIA Nemotron Personas. We present data on persona conversations and stability, discovering that personas embody background information well, yet exhibit hallucinations and drift behavior in exchange, as well as multilingual instability. We release conversation data publicly for AI alignment in multilingual, multicultural, and persona-based prompt scenarios.