A puzzle-solving agent using the ReAct pattern commits to one Thought, takes one Action, reads the Observation, and moves forward to its next Thought in a single unbroken line -- it never goes back to reconsider an earlier step once it has moved on. For puzzles where an early move can look reasonable but turns out to be a dead end many steps later, a different framework instead generates several candidate 'thoughts' at each step, evaluates how promising each one looks, and uses a search procedure such as breadth-first or depth-first search to explore multiple branches and backtrack away from ones that stop looking promising. What is the defining difference this second framework introduces compared to ReAct's approach?
- It removes the need for the model to produce any intermediate reasoning text at all, replacing thoughts entirely with a single numeric score per step
- It requires every branch explored to be executed as a real action against an external tool before it can be evaluated, unlike ReAct which never calls any tools
- It guarantees that the globally best solution will always be found on the first attempt, regardless of how the evaluation heuristic scores each candidate thought
- It reframes problem solving as a search over a tree of intermediate reasoning steps, generating multiple candidate thoughts at a branching point, evaluating their promise heuristically, and backtracking away from weak branches, instead of committing irreversibly to one linear Thought/Action/Observation sequence the way ReAct does
Why D? And why not the others?
Correct answer: D. It reframes problem solving as a search over a tree of intermediate reasoning steps, generating multiple candidate thoughts at a branching point, evaluating their promise heuristically, and backtracking away from weak branches, instead of committing irreversibly to one linear Thought/Action/Observation sequence the way ReAct does
Tree of Thoughts generalizes chain-of-thought-style approaches by treating problem solving as a search over a tree whose nodes are intermediate 'thoughts', generating multiple candidate thoughts at a decision point, using the model itself to heuristically judge how promising each one looks, and applying a standard search procedure like breadth-first or depth-first search to explore and backtrack across branches -- exactly the capability ReAct's single committed Thought/Action/Observation line lacks, since ReAct never revisits or abandons an earlier step once it has moved past it. The option about removing intermediate reasoning text entirely is wrong: the 'thoughts' being searched over are themselves coherent units of reasoning text, not a bare numeric score replacing them. The option requiring every branch to call a real external tool is wrong: this search happens over the model's own generated reasoning candidates and evaluations, and does not depend on tool calls the way agent-environment interaction loops do. The option guaranteeing the globally best solution on the first attempt is wrong: a heuristic evaluation can still misjudge a branch's promise, so the search improves the chance of finding a good solution without guaranteeing the optimal one.
Source: Yao et al., 'Tree of Thoughts: Deliberate Problem Solving with Large Language Models' (2023), arXiv:2305.10601 -- frames problem solving as search over a tree of thoughts using BFS/DFS with heuristic evaluation, contrasted with left-to-right token-level decision-making