New technique reveals how language models process cultural knowledge internally

Researchers have developed a new method to peer inside the inner workings of large language models and understand how they process cultural information—a critical challenge as these systems are deployed globally across societies with distinct values and traditions.

Understanding culture inside the model#

Most prior research on cultural bias in language models has examined what they say: analyzing outputs and responses to cultural questions. But this approach misses a crucial piece of the puzzle—what's actually happening beneath the surface in the model's learned representations. The new technique, called CulTrace, uses mechanistic interpretability methods to trace how cultural knowledge flows through a model's layers during reasoning tasks.

The staged reasoning process#

CulTrace reveals that models follow a consistent pattern when answering culture-related questions. The models first engage with the subject area or domain mentioned in the question, then identify which culture is relevant, and finally zero in on a specific answer. This staged trajectory suggests that cultural reasoning isn't monolithic but unfolds in distinct phases across the network.

Imbalances and blind spots#

A key finding emerges from the analysis: cultural reasoning in these models is uneven. Models take longer to pinpoint the relevant culture when answering questions, and they tend to confuse cultures that are underrepresented in their training data. This imbalance has important implications for real-world deployment, as it suggests that language models may be less reliable when handling knowledge about non-dominant cultures or minority groups.

The research, presented in arXiv paper 2508.08879, opens a new window into how to diagnose and potentially correct cultural reasoning failures in AI systems.