Researchers from the University of Tel Aviv have proposed a method to predict whether incorporating external information into question-answering systems will improve their performance. They evaluated various prediction methods based on retrieval signals, answer characteristics, and semantic consistency between generated responses and retrieved passages. Their approach significantly narrows the gap in performance between computationally demanding post-generation methods and pre-generation methods that use retrieval. The team developed a selective retrieval framework that dynamically chooses whether to use retrieval or not for each question. Experimental results show that selectively applying retrieval augmentation can lead to better answer quality compared to using retrieval for all queries. --- Why it matters: This research matters because it provides insights into how to optimize question-answering systems, which are increasingly used in applications such as customer support and information retrieval. By developing a method to predict when retrieval is beneficial, the team's approach can help improve the efficiency and accuracy of these systems. Source: https://arxiv.org/abs/2604.07985