The Continuous Thought Machine (CTM) is a recurrent network that makes predictions by synchronizing its neurons, according to a new arXiv preprint. Rather than processing inputs through a fixed feedforward pass, the model attends to inputs and reaches decisions through a dynamic synchronization process. The authors describe this as yielding emergent thought processes, though the abstract's details on those processes are cut off in the available text.
The paper, titled "Riding the Fixed-Point Curve," applies CTM specifically to image classification. The abstract notes that the approach works on tasks spanning many domains, suggesting the mechanism is not limited to vision. However, because this is a single preprint and the abstract is truncated, the full scope of the claimed results and how they compare to other methods is not yet clear from this source alone.