This work explores a language model whose internal graph settles through repeated local updates. Nodes make predictions, receive signals from neighboring parts of the graph, and adjust how strongly those signals pass through.
The experiments vary recurrent feedback, state persistence, and constraints on update size. They also examine the connection between continuous internal representations and discrete token predictions.
The main questions concern stability, specialization between nodes, and useful learning over sequences. The project has several framework and implementation variants. Results from one configuration should not be treated as evidence for the whole approach.