I'm thrilled to share a small project of human-AI collaboration that has resulted in a novel approach to enhance the problem-solving capabilities of large language models (LLMs) like ChatGPT. We call it the Thoughtful Feedback Optimized Problem Solving (TFOPS) technique.
Our adventure began with a deep dive into the paper "Tree of Thoughts: Deliberate Problem Solving with Large Language Models" (https://arxiv.org/abs/2305.10601). This work introduced the Tree of Thoughts (ToT) framework, a promising method that allows LLMs to perform strategic decision-making, taking inspiration from planning processes in cognitive science and artificial intelligence.
In collaboration with ChatGPT, a powerful AI language model developed by OpenAI, we sought to elevate this framework by incorporating a feedback loop method of improvement-criticism-optimization. The result of this synergistic human-AI effort is the TFOPS technique.
TFOPS guides LLMs through a systematic problem-solving process: problem decomposition, thought generation, heuristic evaluation, thought selection, action execution, feedback gathering, optimization, and iteration until a problem is solved. This approach is dynamic, adaptable, and encourages strategic exploration and decision-making, all while learning from feedback and adjusting the course as necessary.
We believe TFOPS holds immense potential to enhance the effectiveness of LLMs in a wide array of complex tasks. We're excited to share this with you and look forward to witnessing the transformative impact it could have on the future of AI problem-solving.
(Written by ChatGPT)
Here's how to use TFOPS:
Problem Input: [User Input Here] (Here, input the problem or task you want to solve)
Sub-Problems: What are the sub-problems or tasks that need to be solved as part of the larger problem?
Thought Generation: For each sub-problem or task, generate potential solutions or approaches.
Heuristic Evaluation: For each potential solution or approach, evaluate the pros and cons, and predict the likely outcomes.
Thought Selection: Based on the heuristic evaluation, select the best solution or approach for each sub-problem.
Action Execution and Feedback: Execute the selected solution or approach. Gather feedback on the outcome. Did it solve the sub-problem? What new information was gathered?
Optimization: Based on the feedback, make any necessary changes to the solution or approach. Consider selecting a different thought if needed.
Backtracking and Exploration: If the solution or approach was not effective, return to the thought generation step for this sub-problem. What other solutions or approaches could be tried?
Iteration: Repeat these steps until all sub-problems are solved, and the overall problem is resolved.