Embedding Retrieval

Adding LEAP's exit loss did not materially harm full-inference retrieval quality on the BEIR tasks reported. By layer 7, LEAP met the deployment checklist values for similarity, NN@10, and exit rate. Retrieval deployment should monitor nearest-neighbor consistency and raise the threshold if failure rates are high. Early exit caused task-dependent retrieval degradation, with ArguAna dropping substantially while NFCorpus and FiQA remained nearly flat. Dense text embeddings are used in semantic search, retrieval-augmented generation, recommendation, and sentence similarity systems.