LLM Task Planning With Dependency-Based Subtask Execution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing large language models (LLMs) face challenges in multi-step reasoning tasks due to calculation errors, missing-step errors, and semantic misunderstanding errors, particularly in Zero-shot-CoT prompting.
Innovation Solution
Implementing a Plan-and-Solve (PS) prompting technique that decomposes computing tasks into smaller subtasks and executes them based on dependencies, enhanced with detailed instructions to improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If LLMs execute complex multi-step reasoning tasks directly, then task completion capability is improved, but calculation errors and missing-step errors increase
Solution Approach 1:
The patent segments complex multi-step reasoning tasks into smaller, manageable subtasks that are executed sequentially. Each subtask is processed independently with intermediate results stored and verified, reducing the cognitive load on the LLM and minimizing calculation errors while maintaining overall task completion capability.
2Productivity
If LLMs execute complex multi-step reasoning tasks directly, then task completion capability is improved, but semantic misunderstanding errors increase
Solution Approach 1:
The patent implements preliminary action by having the LLM generate a detailed plan and break down the complex task into subtasks before executing them. This preliminary planning phase allows the system to identify potential semantic misunderstandings early, verify task requirements, and structure the execution flow to prevent errors propagation.
3Reliability
If detailed instructions and task decomposition are implemented, then calculation accuracy is improved, but system complexity increases
Solution Approach 1:
The patent applies self-service by enabling the LLM to automatically generate its own task decomposition, execution plan, and verification steps without requiring external orchestration. The system self-manages the complexity of multi-step reasoning by internally structuring subtasks and dependencies, reducing the apparent system complexity while maintaining high calculation accuracy.
Data Source
AI summary
Disclosed are a system, method and apparatus to define computing tasks for servicing an electronic prompt. Responsive to a first prompt, a second prompt may be submitted to one or more generative neural network models. The second prompts may be based, at least in part, on the first prompt, and may specify a plurality of computing tools for use in constructing a requested response. The second prompt may request an identification of tasks to be executed based, at least in part, on at least some of the plurality of computing tools and based, at least in part, on execution dependencies between and/or among the identified tasks.