AI Reasoning Model Question Decomposition for Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current language model-based reasoning methods struggle to accurately understand natural language questions and distinguish between errors in question understanding and inference processes, leading to difficulties in improving question answering performance and explaining reasoning operations.
Innovation Solution
The proposed method involves summarizing natural language questions into a semi-code-style summary, decomposing the questions into sub-questions, and solving these sub-questions to improve understanding and accuracy in question answering. This method utilizes a device with a processor and memory to execute these operations, enhancing the language model's ability to correctly understand questions and perform multi-step reasoning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If natural language questions are directly input to language models for reasoning, then the process is simple and fast, but the understanding accuracy and reasoning clarity deteriorate
Solution Approach 1:
The patent applies segmentation by dividing complex natural language questions into multiple sub-questions with specific roles (e.g., known conditions, target variables, intermediate steps). This breakdown improves understanding accuracy by making each component clearer while managing complexity through structured organization of the divided elements.
2Reliability
If complex reasoning operations are performed without decomposition, then the process remains unified, but error identification and reasoning clarity worsen
Solution Approach 1:
The patent segments reasoning operations into distinct sub-questions, each handling specific aspects of the problem. This segmentation improves reliability by ensuring thorough analysis of each component while making error detection easier through localized verification points in the segmented structure.
Solution Approach 2:
The patent implements feedback mechanisms where each sub-question's result is verified and used to inform subsequent reasoning steps. This feedback loop enhances reliability by catching errors early and provides clear measurement points for detecting where reasoning deviates from correctness.
3Productivity
If direct inference is performed on natural language questions, then the process is efficient, but the distinguishability between understanding errors and inference errors worsens
Solution Approach 1:
The patent segments the question processing into distinct phases: understanding sub-questions (extracting known conditions and target variables) and inference sub-questions (performing calculations). This segmentation maintains efficiency by streamlining each phase while improving error type distinguishability through clear phase boundaries and structured output formats.
Data Source
AI summary
Provided are a device and method for artificial intelligence (AI) reasoning based on a language model. The method for AI reasoning based on a language model includes generating a question core summary by inputting a natural language question to a language model, decomposing the natural language question into sub-questions using a reasoning model on the basis of the question core summary and the natural language question, obtaining an answer to the natural language question by solving the sub-questions and synthesizing solutions of the sub-questions.