Researchers at UC Davis and Virginia Tech have studied how large language models learn and forget concepts during continuous training. They found that concept circuits in the models provide a signal of learning and forgetting, with an early increase followed by a gradual decrease and stabilization. Concepts with larger learning gains tend to exhibit greater forgetting under subsequent training, while semantically similar concepts induce stronger interference than weakly related ones. --- Why it matters: This study matters because it provides insights into how large language models learn and forget concepts, which can inform the development of more efficient and effective training strategies for these models. This is particularly relevant in applications where concept learning is critical, such as natural language processing and question-answering systems. Source: https://arxiv.org/abs/2601.03570