Researchers have identified a mechanism behind language identity in multilingual transformers. They introduced Language Identity Head Ablation (LIHA), which zeros out each attention head individually and measures the resulting language switch rate. The study found that a small set of 'first-token broadcaster' heads, led by L6H1, attend persistently to the first prompt token and propagate its language signal throughout generation. The researchers also found that instruction tuning reorganizes language identity circuits toward early-layer localization. --- Why it matters: This research matters because it provides direct causal evidence for how instruction tuning affects language identity in multilingual transformers. Understanding this mechanism can help improve the robustness of these models, which is crucial for applications like machine translation and text summarization. Source: https://arxiv.org/abs/2606.22361