ACE-Align: Attribute Causal Effect Alignment for Cultural Values under Varying Persona Granularities

TL;DR

ACE-Align addresses the challenge of aligning LLMs to diverse cultural values under varying persona granularity by explicitly modeling attribute-level causal effects of demographics. It defines a formal causal framework with a DAG, a binary treatment A, context Z, and a latent mediator E, and optimizes a dual objective that aligns $\mathrm{CE}_{\text{LLM}}$ with $\mathrm{CE}_{\text{data}}$ via a $d_{\mathrm{CDF}}$ distance while anchoring model outputs to survey modes. Training at the finest granularity ($G=4$) enables strong covariate control and compositional generalization to novel attribute combinations, leading to improved cultural alignment across 14 countries and across all persona granularities, with Africa showing the largest gains and geographic gaps narrowing from $9.81$ to $4.92$ points. The approach reduces stereotyping and erasure by correcting directional misalignments of attribute effects, and demonstrates substantial gains over baselines such as a base LLM, Anthropological Prompting, and CultureLLM SFT. The work provides a pathway to more equitable, data-grounded cultural alignment for multilingual, globally deployed LLMs, with code available to facilitate reproducibility.

Abstract

Ensuring that large language models (LLMs) respect diverse cultural values is crucial for social equity. However, existing approaches often treat cultural groups as homogeneous and overlook within-group heterogeneity induced by intersecting demographic attributes, leading to unstable behavior under varying persona granularity. We propose ACE-Align (Attribute Causal Effect Alignment), a causal-effect framework that aligns how specific demographic attributes shift different cultural values, rather than treating each culture as a homogeneous group. We evaluate ACE-Align across 14 countries spanning five continents, with personas specified by subsets of four attributes (gender, education, residence, and marital status) and granularity instantiated by the number of specified attributes. Across all persona granularities, ACE-Align consistently outperforms baselines. Moreover, it improves geographic equity by reducing the average alignment gap between high-resource and low-resource regions from 9.81 to 4.92 points, while Africa shows the largest average gain (+8.48 points). Code is available at https://github.com/Wells-Luo/ACE-Align.

Figures (18)

• Figure 1: Within a single culture, personas at different granularities exhibit heterogeneous values, which current LLMs fail to consistently capture.
• Figure 2: The ACE-Align framework. We optimize cultural alignment at the finest granularity ($G=4$) to ensure maximum covariate control and identify pure attribute causal effects. By aligning the model’s response shifts ($\mathrm{CE}_{\mathrm{LLM}}$) with survey-based dynamics ($\mathrm{CE}_{\mathrm{data}}$) via $\Delta$CDF distance, the model learns robust mechanisms that maintain consistency across persona granularities.
• Figure 3: Alignment across resource tiers and persona granularities. (a) Baseline alignment scores over 14 countries reveal a pronounced tier gap (Tier 1, Tier 2, Tier 3 Africa). (b) Gains from ACE-Align over the base model are larger for lower tiers and persist across $G$. (c) ACE-Align reduces the tier gap, indicating cross tier convergence.
• Figure 4: Holdout group 1
• Figure 5: Holdout group 2