Matthias Kern

Extremely excited to announce Promptbreeder 🌱, a self-referential self-improving system that can automatically evolve effective domain-specific prompts in a given domain. Prompts evolved with Promptbreeder outperform Chain-of-Thought and Plan-and-Solve Prompting on a range of arithmetic and commonsense reasoning benchmarks. Creating self-referential self-improving systems is a holy grail of AI research, but prior self-referential approaches rely on costly model parameter updates—it's unclear how to scale them to the vast number of parameters in modern LLMs, let alone how to make them work with LLMs whose parameters are hidden behind APIs. Instead, in Promptbreeder the substrate for self-improvement is natural language, which allows us to modify mutation operators using an LLM in a self-referential way. Thus, Promptbreeder not only improves prompts, but over multiple evolutionary generations it also improves the way it is improving prompts. I keep on coming back to Figure 4 in the original Chain-of-Thought Prompting paper: as the LLM gets bigger and better, the gains from prompt strategies increase significantly. I believe the future is going to be wild as we will see increasingly open-ended self-referential self-improvement systems that continue to scale with ever larger and more capable LLMs. Promptbreeder is an important stepping stone in this direction. Really enjoyed this collaboration with amazing co-workers @chrisantha_f @2ne1 @hmichalewski @sindero arXiv: arxiv.org/abs/2309.16797