Researchers propose a new approach to improve the performance of large language models in query expansion and retrieval tasks. They introduce Document Embedding Preservation Tuning (DEPT), which keeps document embeddings stable while allowing for end-to-end training of the model. This is achieved by passing retrieval gradients through straight-through decoding into the generator, while keeping tuned document embeddings close to their initial values. The authors claim that DEPT improves average retrieval quality over various baselines on five datasets in the BEIR benchmark. --- Why it matters: This matters because it addresses a common challenge in large language models: how to keep generated expansions aligned with the retrieval loss that judges them. By stabilizing document embeddings, DEPT enables more effective query expansion and retrieval, which is crucial for applications such as search engines and question-answering systems. Source: https://arxiv.org/abs/2608.17632