I like the branching idea, but I’m not a big fan of inserting “think tokens”. It sort of goes against my ML philosophy, which is to stay on (or close to) the na

I like the branching idea, but I’m not a big fan of inserting “think tokens”. It sort of goes against my ML philosophy, which is to stay on (or close to) the narrow mathematically sound path. So I’d be interested to see how this compares to the mathematically sound approach of MCTS for the highest probability completion (which is not necessarily the same as the greedy / argmax search for the same).