Researchers have proposed a method called CASE to improve the accuracy of large language models by selecting the best answer based on hidden-state signals. The method uses a linear gate to rank candidates and measures its performance with 'decodability', a metric that predicts when hidden-state selection beats majority voting. Experiments show that decodability can predict the accuracy gain of selection over voting, and CASE improves over voting by up to 19 points in certain scenarios. The proposed criterion is leakage-free and transferable across domains. --- Why it matters: This matters because it provides a practical way for engineers to choose between learned selection and majority voting in large language models, which can lead to improved accuracy on difficult questions. Source: https://arxiv.org/abs/2608.17124