Private AI Model Guidance Using Public Reasoning Intermediaries
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI models generate incorrect data due to immature reasoning and privacy concerns when processing query documents, especially when utilizing very large language models (VLLMs) for domain-specific tasks, which are expensive, slow, and have unverified privacy practices.
Innovation Solution
A two-step process involving a very large language model (VLLM) to generate general guidance, updated with domain-specific information, iteratively guiding smaller AI models to answer reasoning questions, ensuring privacy by filtering documents and utilizing smaller models for faster, cheaper processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If very large language models (VLLMs) are used to answer reasoning questions, then the accuracy and reasoning capability are improved, but the processing speed and cost increase
Solution Approach 1:
The patent segments the AI processing task into two distinct stages: a VLLM-based reasoning stage for generating guidance principles, and a smaller AI model stage for executing specific answers. This segmentation allows the system to leverage the superior reasoning capability of VLLMs without relying on them for all processing tasks, thereby improving overall processing speed while maintaining high accuracy.
Solution Approach 2:
The patent introduces an intermediary mechanism where the VLLM generates general guidance principles that serve as mediators for smaller AI models. Instead of directly using VLLMs for all queries, the system uses these guidance principles as intermediaries to guide smaller, faster models in generating specific answers, thus resolving the speed-accuracy tradeoff.
2Reliability
If very large language models (VLLMs) are used to process query documents, then the reasoning capability is improved, but the cost and processing time increase
Solution Approach 1:
The patent divides the processing workflow into two segments: first using VLLMs to generate general guidance principles from query documents, then using smaller AI models to process specific questions based on these principles. This segmentation significantly reduces the time VLLMs need to process each individual query while maintaining high reasoning capability.
Solution Approach 2:
The patent applies preliminary action by having the VLLM perform the complex task of generating general guidance principles before the actual question-answering process. This preliminary generation of guidance principles allows subsequent smaller models to operate much faster, as they only need to follow pre-established guidance rather than reasoning from scratch.
3Ease of operation
If public AI models are used for domain-specific tasks, then the accessibility is improved, but the privacy and security concerns increase
Solution Approach 1:
The patent segments the system into a public VLLM component for generating general guidance and a private smaller model component for processing sensitive domain-specific questions. This segmentation allows the system to leverage the accessibility of public models while maintaining data privacy by processing actual queries through private, locally-deployed smaller models.
Solution Approach 2:
The patent uses generated guidance principles as intermediaries that bridge public and private model components. The VLLM generates these guidance principles from public access, which then serve as instructions for private smaller models to process sensitive data locally, thus maintaining both accessibility and privacy.
4Productivity
If smaller AI models are used for processing, then the cost and speed are improved, but the accuracy and reasoning capability decrease
Solution Approach 1:
The patent applies preliminary action by having the VLLM perform complex reasoning tasks and generate detailed guidance principles before the smaller models execute specific answers. This preliminary reasoning work compensates for the limited capabilities of smaller models, enabling them to achieve high accuracy while maintaining processing efficiency.
Solution Approach 2:
The patent implements a feedback mechanism where the VLLM's generated guidance principles are fed back to smaller AI models as instructions. This feedback loop allows smaller models to leverage the reasoning capabilities of VLLMs indirectly, improving their own accuracy while maintaining their speed and cost advantages.
Data Source
AI summary
Systems and methods for guiding private artificial intelligence models with public solutions. A very large language model (VLLM) can be iteratively queried with an instruction code including public entities with associated public documents to generate public solutions. Rationale features can be extracted from the public solutions with the VLLM. The instruction code can be updated by combining an input query about public entities, the public solutions with the rationale features, text from reference chunks, and an input query about a single private entity, following a pre-determined template, to yield a private instruction code about a single private entity. The private instruction code can be answered with private large language models (PLLM) to obtain private answers for performing downstream tasks.