← Back to concept explainers

Concept Fable

Graph of Thoughts

Conclusion first:Graph of Thoughts organizes intermediate ideas in a graph structure, allowing reasoning nodes to branch, merge, backtrack and be reused. Compared with the tree search of Tree of Thoughts, it is better suited to tasks that need to combine multiple intermediate results, share subproblems or repeatedly rewrite plans. The cost is more complex state management and evaluation, requiring explicit nodes, edges, merge rules and search budget.

Graph of Thoughts · Reasoning

Read the story first. There is no rush to give a definition here; let the question grow on its own first.

Fable

The canal builder discovered the branching paths could merge again

When building irrigation channels on the plains, the craftsmen first planned them like tree branches: one main channel branching into several sub-channels, each sub-channel branching into smaller ditches. This made drawing simple and construction easy.

Later they encountered complex farmland. The drainage plan from the east could help the west store water, and the slope tested in the south could be reused in the north. If each branch canal were unrelated, much experience would be wasted through repetition.

The supervisor first had everyone draw more trees. Each field had its own solution tree, the drawings piled up thicker and thicker, the same problem appeared repeatedly on different trees, and failed detours were walked again and again.

The young surveyor proposed drawing the ditches as a network. An idea can branch or merge; a piece of computation can be referenced by multiple solutions; useful parts of a failed path can also connect elsewhere.

The old craftsman worried the network would become a mess. So the surveyor marked each node with its source, purpose and connection conditions, and only truly reusable waterways were connected together. Not every line can be linked; the connection itself must have a reason.

They began merging multiple local solutions. A low spot found by one branch helped another branch avoid standing water; measurement values left by a failed solution were connected into a new solution and continued to be used.

Before completion, the northern plan failed as a whole, but one method within it for measuring the mud layer was excellent. Under the old tree diagram, this would have been discarded along with the entire branch; under the new network diagram, it was connected to the western plan, saving three days of re-measurement.

The supervisor finally saw that complex problems are not always like a tree that can only grow downward. Better answers often come from filtering, revisiting and recombining several intermediate ideas. The diagram is harder to manage, but it saves a lot of repetition and waste in exploration. The surveyor also set cleanup rules for this network. Useless branches must be closed, expired measurements must be marked red, and contradictory nodes must not be quietly merged together. Otherwise the network becomes a tangled mess. The benefit is not in having many connections, but in preserving reusable intermediate results. Later, when new craftsmen entered, what they first learned was not how to draw more branches, but how to judge whether an intermediate result could be safely borrowed by others. If it can be borrowed, the diagram has value. Later, during the retrospective, everyone wrote this lesson into the daily rules: first look at where on site the result truly changed, then decide what to keep, what to repair and what to restrict.

Reveal

This story is about: a thought graph

Graph of Thoughts organizes intermediate ideas in a graph structure, allowing reasoning nodes to branch, merge, backtrack and be reused. Compared with the tree search of Tree of Thoughts, it is better suited to tasks that need to combine multiple intermediate results, share subproblems or repeatedly rewrite plans. The cost is more complex state management and evaluation, requiring explicit nodes, edges, merge rules and search budget.

What matters is that it is not just a term, but a structural problem that repeatedly appears in real AI / Agent systems. Understanding it lets you judge when to add a model, when to change a process, and when to add governance.

Metaphor mapping

  • Branching canals: Tree of Thoughts
  • Channels drawn as a network: Graph of Thoughts
  • Nodes indicate source and purpose: thought node metadata
  • Branching and merging: branching and merging
  • The failure path leaves behind measurements: reuse intermediate results
  • Merging multiple local solutions: compositional reasoning
  • Networks are harder to manage: the state complexity of graph search
  • Final insight: Graph of Thoughts lets intermediate ideas form a reusable, mergeable reasoning graph

Andy Ren Xu's judgment

The practical value of this concept is that it helps you break "AI features that look smart" into deliverable, verifiable, and governable units of work.

Tree of Thoughtsplanning
京ICP备2026014272号