‘Trace as State’ reruns AI with its earlier reasoning first, a post explains

A post describing the “Trace as State” paper says long-context models often recognize what matters too late: they process the prompt in order, so later insights cannot change how earlier material was read. The proposed fix, as the user describes it, is another pass with the model’s previous reasoning placed first, guiding what it looks for while rereading. The user says the approach won 26 of 27 comparisons on hard long-context tasks.