How AI describes Re2G-Reranker-T Rex

IBM Research's Re2G (Retrieve, Rerank, Generate) is a BART-based sequence-to-sequence model that combines neural initial retrieval with a reranker to improve knowledge-intensive generation tasks. The system uses a novel knowledge distillation approach to train retrieval, reranking, and generation end-to-end, achieving 9% to 34% relative gains over prior state-of-the-art on the KILT benchmark for tasks like slot filling, question answering, fact checking, and dialog.