Researchers have developed a method called Latent Space Refusal Anchoring (LSR-Anchoring) to help AI models refuse harmful requests in low-resource African languages, such as Yoruba and Igbo. This approach extracts the refusal direction from English prompts and applies it to the model's residual stream at inference time, without requiring retraining or labeled data. The method was tested on several architectures and showed promising results, with some languages transferring positively but others failing due to a geometric mismatch. --- Why it matters: This matters because current AI models often struggle to refuse harmful requests in low-resource languages, which can have serious consequences. LSR-Anchoring provides a potential solution by enabling models to reject unsafe inputs without requiring extensive retraining or data labeling. Source: https://arxiv.org/abs/2608.18089