Large language models are often trained to reflect a single, Western-centric value system, which can be problematic for global users. A new preprint on arXiv tackles the harder problem of fine-grained value alignment—making models reflect the values of specific subgroups.
The study, titled "Can Persona-Prompted LLMs Emulate Subgroup Values?", explores whether prompting a model with a persona (e.g., a demographic or cultural identity) can steer its outputs toward that subgroup's values. The authors also examine the generalisability of this approach across different subgroups and its fairness—whether it works equally well for all.
Because the abstract is cut off, the specific results are not available in the summary. However, the research addresses a key limitation of current alignment methods, which often assume a single set of values. The findings could inform how LLMs are adapted for diverse cultural contexts.