When I first saw Tree of Thoughts, I asked myself: If language models can reason better by searching, why don't they do it themselves during Chain of Thought? Some possible answers (and a new paper): 🧵