Protoreasoning enables step-by-step reasoning in 1M-parameter transformers

Protoreasoning allows tiny, 1M-parameter models to perform a simple form of Chain of Thought on Dyck language tasks. This method reduces the out-of-distribution generalization gap and enables granular study of reasoning algorithms at scales far below natural-language competence.