Tree of Thoughts

Tree of Thoughts explores several possible next steps at once, scores each one, keeps the most promising few, and repeats, like a search over lines of reasoning. Use it for combinatorial reasoning, multi-step planning, math (Game of 24) and puzzles: anywhere a single straight-line ReAct attempt would commit to a bad path too early. It costs far more than ReAct (5–10×). From Yao et al. 2023, Tree of Thoughts: Deliberate Problem Solving with Large Language Models.

Pattern (BFS beam search)

1. Root is the problem statement.
2. For each level up to max_depth: a. Expand. For every frontier node, the proposer generates branch_factor candidate “thoughts” (next steps toward a solution). b. Evaluate. The evaluator scores each candidate 0–1 (how promising is this branch?). c. Prune. Keep only the top beam_width scored candidates. d. Early exit. If any candidate scores >= solved_threshold, stop early and use that branch.
3. Best of the last level wins. The highest-scoring node of the deepest level reached (the solved node on an early exit) picks the winning chain of thoughts, as in Yao et al.’s BFS; an interior step never beats the steps built on it. By default one more model call (synthesize_final=True) turns that chain into the final answer.

Usage

from loomflow import Agent
from loomflow.architecture import TreeOfThoughts
 
agent = Agent(
    "Solve the puzzle by exploring multiple lines of reasoning.",
    model="claude-opus-4-7",
    architecture=TreeOfThoughts(
        branch_factor=3,        # candidates per node
        beam_width=2,           # keep top-2 each depth
        max_depth=4,
        solved_threshold=0.95,  # early-exit if any candidate hits this
        min_score=0.3,          # drop branches below this
    ),
)
result = await agent.run("Game of 24 with [3, 7, 8, 1].")

Or by string:

agent = Agent("...", model="...", architecture="tree-of-thoughts")

Cost

• LLM calls per depth: roughly frontier_size × (branch_factor + 1) . Propose + evaluate.
• Total: at full depth, O(branch_factor^depth) proposals before pruning, then beam_width × max_depth after.
• Real numbers: branch=3, beam=2, depth=4 → ~30 LLM calls vs ~5 for ReAct. Reserve for tasks where ReAct’s first commit is wrong.

When ToT pays off

• Math, puzzles, planning: Game of 24, crosswords, mini Sudoku, constraint satisfaction.
• Combinatorial generation: Story branches, counterfactual exploration.
• Backtracking-friendly problems: Situations where committing to the wrong first move is hard to recover from in ReAct.