Until now, large language models have relied on human-led retraining to adjust their reasoning and update the parameters that shape their understanding.