During reasoning, language models can exhibit sharp phase transitions where the distribution of potential answers changes dramatically