LangFIR: Discovering Sparse Language-Specific Features from Monolingual Data for Language Steering

TL;DR

LangFIR introduces a novel method to identify sparse, language-specific features in multilingual language models using only monolingual data and random-token filtering, enabling effective language control.

Contribution

This work presents LangFIR, a new approach that discovers language-specific features without requiring multilingual or parallel data, outperforming existing methods in language steering accuracy.

Findings

LangFIR finds highly sparse, language-specific features in residual streams.

Directional ablation of these features increases cross-entropy loss for the target language.

LangFIR achieves superior BLEU scores across multiple models, datasets, and languages.

Abstract

Large language models (LLMs) show strong multilingual capabilities, yet reliably controlling the language of their outputs remains difficult. Representation-level steering addresses this by adding language-specific vectors to model activations at inference time, but identifying language-specific directions in the residual stream often relies on multilingual or parallel data that can be expensive to obtain. Sparse autoencoders (SAEs) decompose residual activations into interpretable, sparse feature directions and offer a natural basis for this search, yet existing SAE-based approaches face the same data constraint. We introduce LangFIR (Language Feature Identification via Random-token Filtering), a method that discovers language-specific SAE features using only a small amount of monolingual data and random-token sequences. Many SAE features consistently activated by target-language inputs do not encode language identity. Random-token sequences surface these language-agnostic features, allowing LangFIR to filter them out and isolate a sparse set of language-specific features. We show that these features are extremely sparse, highly selective for their target language, and causally important: directional ablation increases cross-entropy loss only for the corresponding language. Using these features to construct steering vectors for multilingual generation control, LangFIR achieves the best average accuracy BLEU across three models (Gemma 3 1B, Gemma 3 4B, and Llama 3.1 8B), three datasets, and twelve target languages, outperforming the strongest monolingual baseline by up to and surpassing methods that rely on parallel data. Our results suggest that language identity in multilingual LLMs is localized in a sparse set of feature directions discoverable with monolingual data. Code is available at https://anonymous.4open.science/r/LangFIR-C0F5/.