Researchers have been trying to understand why adversarial training reduces the phenomenon of superposition. Superposition occurs when a model uses multiple features to represent something, but it can also lead to errors. A new study has found that adversarial training works by reducing the number of features used by the model, which in turn reduces superposition. This is based on an analysis inspired by previous work on feature taxonomy. --- Why it matters: This matters because understanding why adversarial training works could help improve its effectiveness and reduce errors in AI models. Source: https://arxiv.org/abs/2608.22155