LLM Output Adaptation Using Provider Evaluation Feedback
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing large language models (LLM) have limited applicability and may not be suitable for specific services like content provision, such as targeted advertising, due to insufficient adaptability and evaluation mechanisms.
Innovation Solution
An information processing apparatus that includes a reception unit to receive evaluations of LLM output and a learning unit to adapt a second learning model based on content provider feedback, allowing for tailored output generation through a relationship learning process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a general large language model is used for content provision services, then the model can handle diverse queries, but the model lacks adaptability to specific service requirements and content provider preferences
Solution Approach 1:
The patent segments the learning process into two distinct models: a first learning model for general output generation and a second learning model for service-specific adaptation. This segmentation allows the system to maintain a general-purpose LLM while adding specialized adaptation capabilities through a separate component that learns service-specific preferences and evaluation criteria from content provider feedback.
Solution Approach 2:
The patent creates a universal system that can serve multiple content providers with different preferences through the second learning model. This model learns to adapt the first learning model's output for different content providers by incorporating their specific evaluation criteria, enabling one system to fulfill multiple specialized functions without requiring separate models for each provider.
2Manufacturing precision
If the first learning model generates output without evaluation feedback, then the generation process is fast, but the output relevance to content provider requirements is insufficient
Solution Approach 1:
The patent implements a feedback mechanism where content providers evaluate the output of the first learning model, and this evaluation feedback is used to train the second learning model. The second learning model then uses this learned feedback to adjust and optimize the first learning model's output, improving relevance while maintaining generation efficiency through the intermediary learning layer.
Solution Approach 2:
The patent performs preliminary learning of service-specific patterns and preferences through the second learning model before actual content generation. By pre-learning the relationship between general outputs and content provider requirements during the training phase, the system can quickly adapt outputs during deployment without requiring real-time evaluation feedback for each generation.
3Reliability
If a single learning model is used for all content providers, then the system is simple to manage, but it cannot capture individual content provider preferences and evaluation criteria
Solution Approach 1:
The patent divides the learning system into two functional segments: a first learning model that handles general language understanding and generation, and a second learning model that specializes in learning content provider-specific preferences and evaluation criteria. This segmentation allows each model to focus on its specific function, improving customization accuracy while keeping individual model complexities manageable.
Solution Approach 2:
The second learning model acts as an intermediary between the first learning model and the content providers. It translates general model outputs into provider-specific customized outputs by learning the mapping between general patterns and provider preferences, without requiring direct complex interactions between the base model and each provider's specific requirements.
Data Source
AI summary
An information processing apparatus according to the present application includes a reception unit and a learning unit. When predetermined input information is input to a first learning model learned to generate, as output information, an answer to a question input as input information, the reception unit receives information concerning evaluation for output information output by the first learning model and corresponding to the predetermined input information. The learning unit causes, based on content provision information received from a content provider that provides content in a predetermined service, a second learning model, which is another learning model different from the first learning model, to learn, for each content provider, a relationship between a change policy for the learning model to generate desired output information in the case in which the output information output by the first learning model is changed and the evaluation.