Summary
Large language models that work with many languages sometimes show stereotypes, but it’s not clear how or where these biases appear inside the models. The authors studied three different AI models and used several techniques to find where stereotype-related information lives within them and how it influences what the models say. They found that some parts of the model reveal stereotypes early on, while others show effects later, and this timing differs by model. They also discovered that only a small fraction of features work the same way across languages, and none behave the same way across different social categories. This means measuring bias inside AI is complex and needs different approaches to understand fully.
Multilingual large language modelsStereotype biasLinear probingAttribution patchingSparse autoencodersFeature ablationResidual stream featuresCross-lingual effects
Authors
Ariun-Erdene Tumurchuluun, Yusser Al Ghussin, Pinzhen Chen, Josef van Genabith, Koel Dutta Chowdhury
Abstract
Multilingual LLMs show stereotype-related behavior that varies across languages, but behavioral scores do not show where the relevant information is represented or how it affects the output. To investigate these internal mechanisms, we compare linear probing, attribution patching, sparse autoencoders (SAEs) and feature ablation in Llama-3.1-8B, Qwen3-8B, and Gemma-2-9B. Probe performance peaks substantially earlier than attribution in all three models, with a separation of 36-53% of model depth. Retained Llama-Scope features often match the social category on which they were selected and form recurring semantic families, but their lexical alignment and ablation effects vary across SAE suites. Only 6-18% of evaluated residual-stream features have language-agnostic effects under our criterion, and none are category-agnostic. Language-agnostic features have larger mean ablation effects in Llama-Scope, but this pattern does not repeat in the other SAE suites. Decodability, output influence, and cross-lingual ablation effects therefore need to be measured separately.