Latent reasoning models attempt to think in a continuous space instead of generating explicit chain-of-thought tokens. According to the abstract for arXiv:2610.17189, the usual way to do this is to replace discrete tokens with soft tokens. But the abstract is blunt about the current state of the art: these approaches typically need multi-stage training and, in some cases, auxiliary modules, and they still underperform.
The paper, titled "Soft Teacher Forcing," appears to respond to that gap by proposing a training strategy. The abstract does not include results or implementation details, so it is not possible to say how well the method works or how it differs from existing training schemes. What the source does establish is the problem: soft-token latent reasoning is heavy and underdelivers, leaving room for alternative approaches.