Tag
This paper introduces the Bicameral Model, which couples two frozen language models through a trainable neural interface on their intermediate hidden states to enable continuous, concurrent coordination without serialized text exchanges. The approach demonstrates significant improvements in arithmetic and logic tasks by allowing an auxiliary model to operate tools in parallel with a primary model.