A growing number of research teams are creating self-improving systems that use LLMs to automatically generate hypotheses and reflect on the results.