Tree of Problems: Improving structured problem solving with compositionality
Tree of Problems (ToP) breaks complex LLM tasks into identical subtasks, solving them like nested Russian dolls
Tree of Problems (ToP) breaks complex LLM tasks into identical subtasks, solving them like nested Russian dolls
Turns massive problems into bite-sized copies.
Original Problem 🤔:
LLMs struggle with complex reasoning tasks that require breaking down into simpler subtasks. Existing methods like Chain-of-Thought (CoT) and Tree of Thoughts (ToT) can be overly complex or fail out-of-domain.
Solution in this Paper 🛠️:
Introduces Tree of Problems (ToP), a framework that decomposes complex tasks into identical subtasks using a tree structure.
Decomposer: Splits the main problem into smaller instances.
Solver: Solves atomic subproblems using LLMs with task-specific prompts.
Merger: Combines solutions from subproblems to solve higher-level nodes recursively.
Efficient for both canonical and sequential tasks, enhancing problem-solving capabilities.
Key Insights from this Paper 💡:
ToP simplifies problem-solving by focusing on analogous subproblems.
Outperforms CoT, ToT, and GoT in structured tasks.
Enhances LLMs' generalization and accuracy on complex tasks.
Results 📊:
Sorting: ToP achieved 68% accuracy vs. 28% for GoT.
Last Letter Concatenation: 99% accuracy on four-word lists, surpassing CoT.
Sequential tasks like Coin Flip showed near-perfect accuracy with ToP.
🌳 How the Tree of Problems framework works
ToP builds a hierarchical tree structure where each node is a subproblem similar to the main task. The process involves:
Decomposer: Splits the main problem into smaller instances.
Solver: Uses LLMs with specific prompts to solve these atomic subproblems.
Merger: Combines solutions from subproblems to solve higher-level nodes, ultimately solving the main problem.
