New methods steer recurrent reasoners at inference time

Two arXiv papers propose controlling recurrent-depth reasoners at test time. One uses readout feedback to steer latent states; the other shows a measurable property of the trained operator predicts whether extra iterations improve or degrade answers.