Researchers have proposed a new approach to test-time reasoning in AI systems. The method, called 'candidate-free control,' eliminates the need for selecting or recombining candidates during inference. Instead, it uses a fresh solve to aggregate answers. In experiments on two mathematics benchmarks, the candidate-free control showed improved accuracy when multiple candidates were correct but decreased accuracy when all candidates were wrong. The results suggest that conditioning on an all-wrong candidate pool can actually lower accuracy compared to a fresh solve. --- Why it matters: This matters because it challenges conventional wisdom in AI test-time reasoning and highlights the importance of considering the context in which answers are generated. It also has implications for the design of efficient inference mechanisms, particularly when dealing with multiple candidates or uncertain outputs. Source: https://arxiv.org/abs/2608.18379