Graph-of-Thought (GoT)

Expert

Model thoughts as graphs to enable arbitrary pathways.

01 The Concept

While Chain-of-Thought (CoTe) models reasoning linearly, and Tree-of-Thought (ToT) branches into alternatives, Graph-of-Thought (GoT) models thoughts as a Directed Acyclic Graph (DAG).

In a graph architecture: - Thoughts can branch out to explore alternatives. - Multiple thoughts can aggregate/merge (e.g. summarizing three different ideas into one solution). - Steps can loop back to refine earlier nodes (self-correction).

GoT is essential for complex logic puzzles, multi-agent debates, sorting algorithms, and document merging tasks.

02 Weak vs. Strong

EX 01Document Merger & Consensus (Python)

[External Controller runs DAG orchestration] # Node 1: Generate analysis of proposal A # Node 2: Generate analysis of proposal B # Node 3: Generate analysis of proposal C # Node 4 (Aggregation): Merge analyses of 1 & 2 to find overlaps # Node 5 (Aggregation): Merge analyses of 2 & 3 to find overlaps # Node 6 (Consensus): Read Node 4 and Node 5 outputs, resolve differences, and write the final roadmap document.

→ Why it works

GoT models multi-faceted consensus and cross-checking as graph nodes. The model doesn't just read everything in one pass, but systematically aggregates and filters ideas.

03 Key Points

01Arbitrary Topologies: Transcend tree-based and linear pathways to model complex cognitive maps.

02Aggregation Nodes: Explicitly design points where the model combines outputs from multiple previous reasoning streams.

03Scoring and Pruning: Evaluate individual thought nodes; discard weak paths and merge strong ones.

04State Tracking: Requires an external controller (Python/JS framework) to route output variables between prompts.

05Feedback Loops: GoT allows a critique node to feed back into a generation node to run iterative refinement.

04 Model-Specific Notes

Claude is excellent at aggregation and consensus tasks due to its strong contextual synthesis capabilities.

05 For Your Role

It's like a brainstorming session: team members brainstorm separately (branching), combine their best ideas (merging), and refine the results (looping).

Next: Contrastive Chain of Thought (CCoT) →