Tropical Reinforcement Learning

Arip Asadulaev, Aladin Djuhera, Karim Salta +3

Reinforcement learning for large language models typically maximizes expected return, adding up the probabilities of all successful trajectories. However, the classical sum formula…