Tree of Thoughts
Overview
Tree of Thoughts studies how language models can solve problems by generating multiple candidate thoughts, organizing them as branches, and evaluating which paths to continue. Approaches include breadth-first and depth-first search, Monte Carlo tree search, and learned or prompted heuristics for scoring, pruning, and allocating computation. The methods are used in settings such as mathematical reasoning, planning, code repair, clinical diagnosis, and multimodal tasks. A central challenge is balancing accuracy and exploration against inference cost: aggressive pruning can miss useful paths, while broader search requires more time and memory.
Latest papers
15 papers0 papers in the last four weeks, with none the four weeks before. 0.0% of all new papers.