Instead of following dynamic situations like concentration games step-by-step, language models use mathematical shortcuts to make predictions. Engineers can control when these workarounds are used to help the systems make better predictions.