Papers for

model interpretability engineers

Papers whose findings have a practical use for this group, as judged from the abstract. Open a paper to read what it means in practice.

Multilingual large language models show language specific stereotype patterns inside

Tracing Stereotypes from Representation to Output in Multilingual LLMs

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.

Tue 8 SeptComputation and LanguageArtificial Intelligence
The gist
Large language models that work in many languages show behaviors related to stereotypes, but it is unclear where these biases live inside the models or how they influence what the model says. The authors studied three such models using various analysis methods to find where stereotype information appears in their internal layers. They found that signals related to stereotypes appear much earlier inside the model than when they actually affect the output, and that these signals vary by language and social category. This shows that checking bias requires looking separately at what the model internally represents and what it actually uses to produce its answers.
Open → 2609.08322v1