CALM: Culturally Self-Aware Language Models

TL;DR

CALM tackles cultural awareness in language models by internalizing culture as a dynamic reasoning dimension rather than fixed background knowledge. It introduces an Abstract Cognitive Space, a Contrastive Window, an Identity Alignment Pool with cross-attention and a culture-aware Mixture-of-Experts, and a Reflective Reasoning Loop to continually calibrate cultural alignment. Across CultureAtlas, UniVaR, CREHate, and EMGSD, CALM outperforms strong baselines and demonstrates robust, cross-cultural generalization; ablations confirm the necessity of each component. By fusing symbolic and pragmatic cultural signals into a cohesive internal identity, CALM enables more culturally sensitive and consistently adaptive reasoning, with potential extensions to time-varying culture and multi-modal data.

Abstract

Cultural awareness in language models is the capacity to understand and adapt to diverse cultural contexts. However, most existing approaches treat culture as static background knowledge, overlooking its dynamic and evolving nature. This limitation reduces their reliability in downstream tasks that demand genuine cultural sensitivity. In this work, we introduce CALM, a novel framework designed to endow language models with cultural self-awareness. CALM disentangles task semantics from explicit cultural concepts and latent cultural signals, shaping them into structured cultural clusters through contrastive learning. These clusters are then aligned via cross-attention to establish fine-grained interactions among related cultural features and are adaptively integrated through a Mixture-of-Experts mechanism along culture-specific dimensions. The resulting unified representation is fused with the model's original knowledge to construct a culturally grounded internal identity state, which is further enhanced through self-prompted reflective learning, enabling continual adaptation and self-correction. Extensive experiments conducted on multiple cross-cultural benchmark datasets demonstrate that CALM consistently outperforms state-of-the-art methods.

Figures (8)

• Figure 1: A case of culturally sensitive food recommendation. Left: AI assistant recommends beef to a user from Varanasi, resulting in discomfort. Right: Recommendations lead to a satisfying outcome.
• Figure 2: Overview of the CALM framework, comprising four stages that progressively embed cultural awareness into internal reasoning and enable self-consistency calibration.
• Figure 3: Left panel: Cultural feature alignment process. Right panel: MoE schematic diagram.
• Figure 4: Log-scale statistics of CALM’s training time, epochs, and sample size across four datasets.
• Figure 5: Proportion of stereotypical responses generated by each model on 1,000 culturally neutral prompts derived from the EMGSD benchmark.