LLM-Generated Curriculum Learning Boosts Blackjack AI Performance by 3.4%, Cutting Training Time by 74%

Researchers propose a framework using Large Language Models to dynamically create training curricula for RL agents, progressively introducing complex actions

Applied to Blackjack with both Tabular Q-Learning and Deep Q-Network (DQN) agents tested on realistic 8-deck simulations

DQN agent's win rate improved from 43.97% to 47.41%, while bust rate dropped from 32.9% to 28.0%

Curriculum-based approach accelerates training by 74%, with full agent training completing faster than standard baseline's evaluation phase alone