Monte Carlo Tree Search training could improve language models step by step

Original: Monte Carlo Tree Search training corrects the model move by move, while current LLM training only tells it whether the whole trajectory worked.

Source: x.com ↗