Doing More with Less: How Cultural Personas Bridge the Cost-Performance Gap in LLMs

Abstract

Cultural personas have emerged as a promising technique for enhancing creativity in Large Language Models, yet their effectiveness across different architectures, domains, and economic contexts remains poorly understood. This study provides a systematic evaluation of three cultural personas—Culture\_5 (innovator), Culture\_Expert (implementer), and Culture\_11 (cultural specialist)—across seven large language models representing both Western (Claude, Gemini, GPT, Llama, Mistral) and Eastern (DeepSeek, Qwen) training regions. (Gemini's unique response format, containing reasoning traces not present in other models, limited direct comparability; results are reported for completeness.) Using 100 problems spanning four domains (Creative, General Knowledge, Instruction, Reasoning), we analyze 31,605 judge ratings of 3,516 model responses using a multi-judge evaluation framework, treating the response (the mean of its nine judge ratings) as the unit of analysis. Inter-rater reliability over the full dataset was fair-to-moderate for all four core creativity dimensions (Fleiss' \(\kappa = 0.31\)--\(0.33\)); we focus on Novelty as the construct central to our research question. Persona effectiveness follows a contingency model rather than a universal rule: effects are highly model-dependent. Claude with Culture\_5 showed a large positive effect on Novelty (\(d = 1.087\)), while Qwen with Culture\_11 showed a medium negative effect (\(d = -0.638\)); 11 of 21 model-persona comparisons reached statistical significance. Domain-specific analysis revealed that Culture\_5 performed best on General Knowledge (\(d = 0.628\)) and Reasoning (\(d = 0.545\)), while Culture\_11 harmed Novelty across all domains (\(d = -0.149\) to \(-0.229\)). Economically, six model-persona combinations achieved significant performance gains at equal or lower cost than their large baselines (40-87\% savings for GPT and Claude; cost parity for DeepSeek). The Efficiency Gain Ratio (EGR) provides a unified metric: GPT combinations achieved exceptional economic value (EGR up to 9.15, where EGR > 5 indicates exceptional value). Training region did not moderate persona effects (interaction \(p = 0.445\); Bayes Factor \(= 0.331\)), suggesting training origin is less important than architectural factors. These findings demonstrate that persona effectiveness is contingent on model architecture and task domain, and establish the economic viability of small models with personas for novelty generation. For practitioners, the path forward is not to seek a single "best" persona, but to match the right persona to the right model for the right task.