New technique steers LLMs' cultural values by manipulating internal activations

Large Language Models (LLMs) often present homogenized cultural perspectives, failing to reflect the diverse values captured by surveys like the World Values Survey. A new paper from Trung Duc Anh Dang and Sarah Masud introduces a framework that moves beyond direct prompting, which typically triggers safety-aligned refusals or neutral responses. Instead, they extract implicit token probabilities across 300 situational dilemmas to map the latent coordinates of LLMs' cultural values. Then, they use activation steering to shift these internal alignments during the forward pass—no retraining needed. This approach provides a computationally efficient way to probe and adjust cultural biases.

The method reveals substantial variation in adaptability across different LLMs. Critically, the researchers uncover a consistent 'latent entanglement' phenomenon: interventions targeting one cultural dimension inadvertently shift others. This suggests cultural values are encoded as coupled structures, limiting the precision of alignment efforts. The work, presented at the ACL 2026 Student Research Workshop, establishes a framework for cultural steering while highlighting the structural complexities of navigating global values with LLMs. It opens doors for more nuanced, region-specific LLM behavior without full retraining.