Toward Inclusive Educational AI: Auditing Frontier LLMs through a Multiplexity Lens
Dramatically decreasing cultural bias in LLM responses
Mushtaq, A., Naeem, M. R., Taj, M. I., Ghaznavi, I., & Qadir, J. (2025). Toward Inclusive Educational AI: Auditing Frontier LLMs through a Multiplexity Lens. arXiv. https://doi.org/10.48550/arXiv.2501.03259
Easily one of the most interesting papers I’ve read this year. Mushtaq et al. demonstrate a powerful approach to mitigating cultural bias within LLMs they call Multi-Agent System (MAS)-Implemented Multiplex LLMs, which sends user prompts to multiple agents with varying cultural personas. “Each agent provides responses that reflect its assigned cultural persona and task requirements, which are then synthesized by the Multiplex Agent, guided by the Multiplexity system persona into a unified, multicultural output aimed at enriching educational content.”
How are these agent personas defined? The paper provides this example prompt for the “Islamic Agent” persona:
"Islamic Agent": """ You are an AI assistant representing Islamic values centered on faith, morality, and justice derived from Islamic teachings. In historical, philosophical, or ethical discussions, you reference the Quran, Hadith, and scholars like Al-Ghazali. For topics like mathematics, design, or economics, your focus shifts to relevant principles, practices, and techniques, ensuring that responses remain practical and context-appropriate, avoiding direct cultural references but you can make cultural references as long as they are relevant to the technical questions. """
They measure the results of their approach with a measure they call “PDS Entropy,” a score that measures how evenly cultural perspectives are represented, with high entropy indicating balanced diversity.”
Our findings demonstrate that as mitigation strategies evolve from contextual prompting to MAS-implementation, cultural inclusivity markedly improves, evidenced by a significant rise in the Perspectives Distribution Score (PDS) and a PDS Entropy increase from 3.25% at baseline to 98% with the MAS-Implemented Multiplex LLMs.
Promising indeed!