Graph-of-Thoughts (GoT) Algorithm

From GM-RKB
(Redirected from GoT Algorithm)
Jump to navigation Jump to search

A Graph-of-Thoughts (GoT) Algorithm is an LLM reasoning algorithm that models the reasoning process of large language models as an arbitrary graph structure (rather than a linear chain or tree).

  • Context:
    • It can provide more flexibility in combining and transforming reasoning paths (more than Chain of Thought or Tree of Thoughts).
    • It can present an innovation is representing inferences as a graph rather than a rigid structure.
    • It can present a graph model that allows feedback loops, merging chains of thoughts, and other operations not possible in Tree of Thoughts.
    • It can (typically) have two main components:
      • a Graph of Operations (GoO): A static, preset graph defined by a HUMAN EXPERT that specifies the high-level reasoning steps and transformations required to solve a particular task.
      • a Graph Reasoning State (GRS): A dynamic structure that stores the execution details - the history of thoughts generated and their states as reasoning proceeds along the GoO graph.
    • A controller module manages execution of the operations laid out in the GoO, coordinating the overall reasoning process. Other modules like prompters and parsers interface with the LLM at each step.
  • See: Tree of Thoughts.


References

2023

  • code_your_own_AI. (2023). “Graph-of-Thoughts (GoT) for AI reasoning Agents". YouTube vlog post
    • SUMMARY:
      • Here is a brief summary of how the Graph of Thoughts (GoT) framework works based on the presentation and paper:
      • GoT consists of several interacting modules - a Prompter, Parser, Scorer, and Controller.
      • The Prompter prepares prompt messages for the LLM. The Parser extracts information from the LLM's responses. The Scorer verifies and evaluates the LLM's replies.
      • The key innovation is the Controller, which coordinates the overall reasoning process. It contains two key components:
        • Graph of Operations (GoO): A static, preset graph defined by a HUMAN EXPERT that specifies the high-level reasoning steps and transformations required to solve a particular task.
        • Graph Reasoning State (GRS): A dynamic structure that stores the execution details - the history of thoughts generated and their states as reasoning proceeds along the GoO graph.
      • The Controller manages the progression of reasoning by executing operations in the GoO graph. The GRS tracks changing parameters as this execution proceeds.
      • The GoO provides an abstract reasoning plan in the form of a graph, while the GRS registers the concrete execution details. Their interaction enables structured yet flexible reasoning that moves beyond linear chains.
      • So, in summary, the Controller drives the reasoning process based on a human-defined GoO graph, with the GRS tracking execution state. The modular architecture coordinates the LLM's inference generation in an interpretable graph framework.

2023