The method formalizes algorithm search as a Markov decision process, enabling credit assignment to specific code components. This removes the need for manually tuning five constants that govern program evolution.

#ReinforcementLearning #LLM #ProgramEvolution