Researchers have found that large language models can recall stored information by scaling input activations with linear projections. This ability is not limited to specific architectures, but rather a general operation in deep learning models. The study defines this phenomenon as 'feature recall' and distinguishes it from feature combination. The findings provide a new conceptual tool for understanding deep learning and suggest empirical directions for mechanistic interpretability research. --- Why it matters: This discovery matters because it highlights the importance of understanding how deep learning models retrieve stored information, which can inform the development of more transparent and explainable AI systems. Source: https://arxiv.org/abs/2608.20970