Graph of Thoughts
A reasoning structure that generalizes Tree of Thoughts to an arbitrary DAG — intermediate thoughts can be combined, refined, or referenced from multiple branches.
Where Chain of Thought is a linear sequence and Tree of Thoughts is a branching search, Graph of Thoughts allows any thought to be an input to any subsequent reasoning step. This enables operations like "merge these two partial solutions" or "refine this earlier thought using this new evidence" that strict tree search cannot express.
Graph of Thoughts is most useful on tasks where partial solutions need to be combined — multi-document synthesis, multi-step proofs, or planning over interacting constraints. The cost is implementation complexity; GoT frameworks are heavier than ToT.
In 2026, GoT-style reasoning is mostly internal to reasoning models (o3, Claude reasoning) rather than something the agent author prompts for explicitly. But the pattern is useful for any agent that needs to reconcile contradictory tool results.
Frequently asked
When does Graph of Thoughts beat Tree of Thoughts?+
When partial solutions need to merge — multi-document synthesis, design-space exploration with shared constraints. For pure search problems, ToT is simpler and usually wins on the same compute.
Do I need a special framework for Graph of Thoughts?+
For research, yes — there are GoT libraries. For production agents, you usually approximate it by storing intermediate thoughts in memory and letting the planner reference them by ID.