Graph of Thoughts

What is it?

Graph of Thoughts treats the generated solution pieces as nodes of a graph and the dependencies between them as edges. A controller runs generation, aggregation, scoring and refinement operations. Writing “think like a graph” in a chat does not set up this execution.

When does it help?

In work where several parts must be processed separately and then combined, or where the same intermediate result must be usable in different branches.

Examples

The situations and responses below are fictional teaching examples; they are not results from a model, tool or benchmark that was actually run.

Simple

Situation

A single list without repeats will be produced from two short lists.

Prompt

A human is the controller; each node is a separate call, and the outputs are kept in a record table. At most 3 calls.
N1 input: "apple, pear, apple". Prompt: Remove duplicates, keep the order of first appearance.
N2 input: "pear, cherry". Same prompt; do not give the N1 output to this call.
N3 input: The real N1 and N2 outputs. Prompt: Merge in the order N1 then N2; remove duplicates.
A human should compare the result with the raw lists. If there is a new word or a missing item, count it as a failure; stop when the budget ends.

Sample output

N1: “apple, pear”; N2: “pear, cherry”; N3: “apple, pear, cherry”. The edges are N1→N3 and N2→N3.

What did we get?

Two separately processed parts met at a single merge point.

Medium

Situation

In an event brief, access and schedule information will be tied to a shared summary.

Prompt

In the controller state, the input, output and source ID of each node should be stored. Sources: D1="Start 14:00, end 16:00"; D2="Entrance has a ramp; lift out of order".
Call N1 extracts the schedule from D1. Call N2 extracts access information from D2. N3 produces a 3-point participant note from the real N1+N2.
Check call N4 receives the raw D1/D2 and N3: show the source for each claim; do not present the broken lift as accessible.
The application should block publishing on any unsourced claim. Stop at 4 calls; no publishing tool.

Sample output

“14:00–16:00 [D1]. There is a ramp at the entrance [D2]. The lift is out of order [D2].” N4 checks that N3 carries only sourced fields.

What did we get?

Besides the graph's merge node, a separate check node was also created.

Hard

Situation

When a conflict appears in the merged draft, only the relevant branch will be reprocessed.

Prompt

Source A: "Hall 20 people". Source B: "Fire safety capacity 16 people". Source C: "Workshop 90 minutes".
Controller: separate extraction nodes for A/B/C; a merge node; a constraint-check node. Add the source and parent node ID to each record. At most 6 model calls in total.
When the check finds the capacity difference, leave the A/B branch to a human decision; do not regenerate the approved duration from C. If the human says "the operational upper limit is 16", merge this decision in as a new node.
If the budget runs out, report the verified duration so far and the open capacity problem. Without a decision, do not average the capacity or pick the higher one.

Sample output

“Duration 90 minutes verified [C]. There is a 20/16 split for capacity; an authoritative decision is pending.” After the decision, the new merge is 16 people / 90 minutes.

What did we get?

The shared state was preserved; the problematic branch did not restart all the work.

Where should you stop?

Tree of Thoughts searches branching candidates; Graph of Thoughts additionally allows nodes to merge and be reused. The existence of a graph does not prove correctness. You need a controller, an evaluation criterion, a call budget and stored real outputs.

Sources

• Graph of Thoughts: Solving Elaborate Problems with Large Language Models (new tab) — Besta, Maciej; Blach, Nils; Kubicek, Ales; Gerstenberger, Robert; Podstawski, Michal; Gianinazzi, Lukas; Gajda, Joanna; Lehmann, Tomasz; Niewiadomski, Hubert; Nyczyk, Piotr; Hoefler, Torsten. 2023-08-18; version read 2024-02-06. Defines organizing solution pieces over a graph through generation, aggregation and refinement operations; the example controllers are simple teaching adaptations of the method. Evidence level: relevant body sections of the original paper.

  Copy source address