Pathway is developing an AI architecture that performs repeated reasoning inside a hidden recurrent state rather than generating a long chain of intermediate text. Its BDH-CQ system builds on a graph of sparsely interacting neuron-like units whose connections retain state during inference.
The company says only about 5 percent of units are active at a time, reducing computation, while the recurrent state can process demonstrations without growing memory in the same way as a transformer’s context and key-value cache. A 150-million-parameter version reportedly set a cost-efficiency record on ARC-AGI-1, a benchmark of visual reasoning tasks. Pathway uses Amazon SageMaker HyperPod to share and scale the computing resources used for development.
These results come from Pathway and AWS, and one benchmark cannot establish a general replacement for transformer models. Reasoning in a latent state also makes direct inspection different from reading a generated chain of thought, despite the company’s claims of architectural interpretability. The work is an early test of whether recurrent computation can deliver useful reasoning with fewer generated tokens and lower inference costs.