Thinking fast, slow, and everywhere in between in humans and language models

How do humans adapt how they reason to varying circumstances? Prior research has argued that reasoning comes in two types: a fast, intuitive type and a slow, deliberate type. Are these the only options, or can people adjust their reasoning continuously by trading off speed and accuracy within individual reasoning steps? We investigate this possibility in an experiment where participants were trained on relationships between local variables in a simple causal model, then asked to make predictions about all pairs of variables. Participants in one condition had a 5-second time limit. We found main effects of time pressure and locality, but only a small interaction in the direction opposite to our hypothesis. We present a process-level model of this phenomenon using early readouts from transformer language models. Our findings are consistent with people reasoning step by step, but accepting a higher error rate at each step under time pressure.