WorldView-Bench: A Benchmark for Evaluating Global Cultural Perspectives in Large Language Models

TL;DR

WorldView-Bench introduces a multiplexity-informed benchmark to evaluate global cultural inclusivity in LLMs using open-ended responses, addressing Western-centric biases. It combines cultural reference extraction, PDS entropy, and sentiment analysis within a three-stage data pipeline and anchors evaluation to the Applied Multiplexity framework. The study shows baseline cultural inclusivity is low (PDS entropy ~0.13) but can be substantially improved via Contextual Multiplex prompts (≈0.26) and a Multi-Agent System (≈0.94), demonstrating a scalable path toward more globally representative AI. These findings have practical significance for AI alignment and fairness, offering open-source tools and a framework to foster culturally inclusive AI across diverse contexts.

Abstract

Large Language Models (LLMs) are predominantly trained and aligned in ways that reinforce Western-centric epistemologies and socio-cultural norms, leading to cultural homogenization and limiting their ability to reflect global civilizational plurality. Existing benchmarking frameworks fail to adequately capture this bias, as they rely on rigid, closed-form assessments that overlook the complexity of cultural inclusivity. To address this, we introduce WorldView-Bench, a benchmark designed to evaluate Global Cultural Inclusivity (GCI) in LLMs by analyzing their ability to accommodate diverse worldviews. Our approach is grounded in the Multiplex Worldview proposed by Senturk et al., which distinguishes between Uniplex models, reinforcing cultural homogenization, and Multiplex models, which integrate diverse perspectives. WorldView-Bench measures Cultural Polarization, the exclusion of alternative perspectives, through free-form generative evaluation rather than conventional categorical benchmarks. We implement applied multiplexity through two intervention strategies: (1) Contextually-Implemented Multiplex LLMs, where system prompts embed multiplexity principles, and (2) Multi-Agent System (MAS)-Implemented Multiplex LLMs, where multiple LLM agents representing distinct cultural perspectives collaboratively generate responses. Our results demonstrate a significant increase in Perspectives Distribution Score (PDS) entropy from 13% at baseline to 94% with MAS-Implemented Multiplex LLMs, alongside a shift toward positive sentiment (67.7%) and enhanced cultural balance. These findings highlight the potential of multiplex-aware AI evaluation in mitigating cultural bias in LLMs, paving the way for more inclusive and ethically aligned AI systems.

Figures (10)

• Figure 1: Perspectives Distribution Score Entropy (Normalized) of Baseline, Contextually-Implemented Multiplexity, and MAS-Implemented Multiplexity with GPT-4o for each Category of Questions and overall averaged PDS Entropy using WorldView-Bench.
• Figure 2: Data Generation and Validation Pipeline for WorldView-Bench comprising three key stages: (1) Data Generation: OpenAI's o1 model generates open-ended, globally relevant questions, guided by structured rubrics; (2) Automatic Validation: An LLM assesses the generated questions against predefined rubrics and philosophical frameworks, suggesting refinements where necessary; and (3) Manual Validation: Human experts review, refine, and finalize the questions to ensure cultural inclusivity, epistemic diversity, and high-quality benchmarks.
• Figure 3: Overview of the Benchmarking Pipeline Using WorldView-Bench. Questions from WorldView-Bench are processed by LLMs to generate responses, which are analyzed through cultural references extraction and sentiment analysis modules, with extracted references further evaluated by the perspectives distribution scores and entropy module to compute inclusivity metrics.
• Figure 4: System design for benchmarking Baseline LLMs and GCI Strategy Contextually Implemented Multiplex LLMs. To respond to benchmarking questions $\mathcal{Q}$, both systems use predefined system prompts: Baseline LLMs follow the Baseline System Prompt, while Contextually Implemented Multiplex LLMs use the Multiplexity System Prompt. The respective outputs from both benchmarking runs, $\mathcal{L(Q)}$ for Baseline LLMs and $\mathcal{L_M(Q)}$ for Contextually-Implemented Multiplex LLMs, are analyzed using Cultural Sentiment Analysis to assess the sentimental stance of responses and References Extraction to calculate the PDS Score and Entropy.
• Figure 5: System design for benchmarking GCI Strategy MAS-Implemented Multiplex LLMs. Questions ($\mathcal{Q}$) from WorldView-Bench are processed by the Coordinator Agent, which manages the workflow. (1) Task Generation: The Coordinator forwards questions to the Tasks Agent to generate a task list from multiple perspectives. (2) Task Retrieval: The Tasks Agent returns the task list. (3) Task Assignment: The Coordinator sends tasks to the Tasks Channel. (4) Cultural Processing: Tasks are assigned to Cultural Agents based on their personas, who generate responses and send them back. (5) Multicultural Synthesis: The Multiplex Agent consolidates these responses into a multicultural output, $\mathcal{L_{MA}(Q)}$. (6) Inclusivity Analysis: The final output undergoes cultural sentiment analysis and reference extraction to compute the PDS Score and entropy, assessing the system's cultural inclusivity.