Proposed: Teaching Large Language Models (LLMs) with Goal-Task Planners

Context

We’re looking at using LLMs and goal-task planners like pyhop to generate innovative, executable plans. This approach combines the creativity of LLMs with the structured planning processes of goal-task planners.

Problem Statement

The main challenges are computational overhead due to the two-step process and potential inaccuracies in the feasibility check.

Proposed Solution

We propose using an LLM for idea generation and a goal-task planner for checking feasibility. This could lead to creative and feasible plans. The accurate plans generated by this model can then be used to further train the LLM.

# Pseudo-code
gpt = GPT()
idea = gpt.generate_idea()

pyhop = Pyhop()
feasible = pyhop.find_plan(state, goal_generator())

if feasible:
    gpt.train(idea)

Benefits

Innovation: Using LLMs for idea generation can lead to creative ideas.
Practicality: Passing the generated ideas through a goal-task planner ensures that the plans are innovative but also feasible and executable.
Continuous Learning: The LLM can learn from the accurate plans, improving its performance over time.

Downsides

Computational Overhead: There might be a computational overhead due to the two-step process.
Accuracy of Feasibility Check: The feasibility check might not always be accurate.

Road Not Taken

An alternative approach could be to run the LLM and the goal-task planner in parallel to address the computational overhead.

Infrequent Use Case

This method may not be suitable for tasks that require immediate results due to the time taken by the two-step process.

In Core and Done by Us

The V-Sekai development team will implement this proposal.

Status

Status: Proposed

Decision Makers

V-Sekai development team