CuMA: Aligning LLMs with Sparse Cultural Values via Demographic-Aware Mixture of Adapters

As Large Language Models (LLMs) serve a global audience, alignment must transition from enforcing universal consensus to respecting cultural pluralism. We demonstrate that dense models, when forced to fit conflicting value distributions, suffer from $\mathbf{\text{Mean}}\mathbf{\text{Collapse}}$, converging to a generic average that fails to represent diverse groups. We attribute this to $\mathbf{\text{Cultural}}\mathbf{\text{Sparsity}}$, where gradient interference prevents dense parameters from spanning distinct cultural modes. To resolve this, we propose $\mathbf{\text{textsc}}\mathbf{\text{CuMA}}$ ($\mathbf{\text{Cu}}$ltural $\mathbf{\text{M}}$ixture of $\mathbf{\text{A}}$dapters), a framework that frames alignment as a $\mathbf{\text{conditional}}\mathbf{\text{capacity}}\mathbf{\text{separation}}$ problem. By incorporating demographic-aware routing, $\text{textsc}CuMA$ internalizes a $\mathit{\text{Latent}}\mathit{\text{Cultural}}\mathit{\text{Topology}}$ to explicitly disentangle conflicting gradients into specialized expert subspaces. Extensive evaluations on WorldValuesBench, Community Alignment, and PRISM demonstrate that $\text{textsc}CuMA$ achieves state-of-the-art performance, significantly outperforming both dense baselines and semantic-only MoEs. Crucially, our analysis confirms that $\text{textsc}CuMA$ effectively mitigates mean collapse, preserving cultural diversity. Our code is available at https://github.com/Throll/CuMA.