Publikation
Tracing Stereotypes from Representation to Output in Multilingual LLMs
Ariun-Erdene Tumurchuluun; Yusser Al Ghussin; Pinzhen Chen; Josef van Genabith; Koel Dutta Chowdhury
In: Hung-Yi Lee; Sunipa Dev; Junyi Jessy Li; Desmond Elliott (Hrsg.). Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing. Conference on Empirical Methods in Natural Language Processing (EMNLP), Budapest, Hungary, Pages 21057-21074, ISBN 979-8-89176-459-0, Association for Computational Linguistics, 2026.
Zusammenfassung
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. 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.