A language-agnostic concept representation causally alters translation across checkpoints

measured in 1 paper

Korner et al. track EuroLLM-1.7B's own pretraining checkpoints, using activation patching to isolate a cross-lingual concept representation and inject it into a translation prompt [korner-etal-2026-when-meanings-meet-investigating-the-emergence-and-quality-of-shared-concept-spaces-during-multilingual-language-model-training] The injected representation causally alters the output translation independent of the source language it was derived from [korner-etal-2026-when-meanings-meet-investigating-the-emergence-and-quality-of-shared-concept-spaces-during-multilingual-language-model-training] Shared concept spaces emerge early in pretraining and continue to refine with more training [korner-etal-2026-when-meanings-meet-investigating-the-emergence-and-quality-of-shared-concept-spaces-during-multilingual-language-model-training] Alignment quality remains language-dependent per a fine-grained manual error analysis [korner-etal-2026-when-meanings-meet-investigating-the-emergence-and-quality-of-shared-concept-spaces-during-multilingual-language-model-training]

Structure

Linear Direction

Context

activation patching used to causally validate that an extracted concept representation is language-agnostic, not merely correlated with translation output, tracking when a shared cross-lingual representation emerges across a model's own training checkpoints, rather than only in a single fully-trained snapshot

Confirmed in models

EuroLLM-1.7B · EuroLLM

Method

Activation patching

Papers

When Meanings Meet: Investigating the Emergence and Quality of Shared Concept Spaces during Multilingual Language Model Training — Körner, Felicia, Müller-Eberstein, Max, Korhonen, Anna, Plank, Barbara2026 · arXiv:2601.22851