GPT-4 can approach a problem by exploring several possible lines of reasoning instead of following only one. A framework called Tree of Thoughts (ToT) makes that possible by pairing the language model with an external process that organizes and evaluates intermediate steps.
Researchers from Princeton University and Google DeepMind developed the approach for language models. Their experiments report substantial gains on some tasks, including a rise in GPT-4’s Game of 24 success rate from 4 percent with chain-of-thought prompting to 74 percent with ToT.
How Tree of Thoughts works
In ToT, a “thought” is a unit of text used as an intermediate step toward solving a problem. Rather than asking the model to produce a single sequence of reasoning, the framework can consider multiple candidate steps and assess which ones appear promising.
That process gives GPT-4 room to change direction. It can look ahead, revisit earlier choices, and compare different paths before deciding how to proceed. The search is guided by the model’s own evaluation and deliberation.
This is a change in how the model is used, not simply a more elaborate prompt. Chain of Thought encourages a model to reason through a task in steps. ToT adds an external module to manage those steps as branches in a search, allowing alternatives to be explored and compared.
Reported gains across different tasks
The Game of 24 result offers a clear example of the difference the framework made in the team’s experiments. With chain-of-thought prompting, GPT-4 solved 4 percent of the tasks. With ToT, its reported success rate was 74 percent.
The researchers also found significant improvements on mini-crossword puzzles and creative writing tasks. Together, these examples suggest that searching across candidate steps can help with both structured puzzles and tasks that are harder to formalize.
Creative writing is a notable case because there may not be one precise procedure for reaching a good result. The framework applies a search process to language model outputs, while the model’s ability to work with open-ended tasks gives that process a broader range of uses.
Search heuristics, without learning
ToT draws a comparison with search heuristics used in systems such as AlphaZero. The similarity is in the use of search to explore possibilities; the difference is that ToT does not learn. It uses GPT-4’s self-evaluation and deliberation to guide the search.
The researchers also connect the approach to Daniel Kahneman’s distinction between System 1 and System 2. In this framing, ToT gives language models a way to spend more effort examining options, rather than relying on a single immediate response.
The framework brings established problem-solving ideas into language model workflows. At the same time, language models may extend those methods to complex tasks that are difficult to express as formal rules, including creative writing.
What the approach could make possible
A separate preprint by a researcher from Theta Labs describes a “Tree of Thought” method that significantly improves performance on Sudoku puzzles. The author points to self-playing techniques associated with AlphaZero as a possible research direction.
One proposed possibility is that a ToT system could develop problem-solving strategies that do not appear in a language model’s training text corpus. The source presents this as a potential direction, rather than an established result.
For now, the reported findings show how adding structured exploration can change a model’s performance on selected tasks. They also point to a broader question: whether models can benefit from examining and revising several possible approaches when a problem does not have an obvious next step.