Autonomous Driving LLM Mediation for User Preference Alignment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing artificial intelligence systems in autonomously driven vehicles lack sufficient training to adapt to diverse user preferences and situational complexities, leading to suboptimal driving behavior.
Innovation Solution
Implementing a dual large language model system, where a second model processes user inputs and queries to refine instructions for a first model, allowing for confirmation and supervision by a policy supervisor to ensure compliance with user desires and regulatory conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a single large language model is used for autonomous driving, then the system structure is simple, but the system cannot sufficiently adapt to diverse user preferences and situational complexities
Solution Approach 1:
The patent divides the single large language model into two separate models: a first large language model for processing driving-related queries and a second large language model for processing user preferences and personalization. This segmentation allows each model to specialize in specific functions, improving overall adaptability while keeping individual model complexities manageable.
Solution Approach 2:
The patent introduces an intermediary module that coordinates between the two large language models and the autonomous driving system. This intermediary manages information flow, resolves conflicts between driving safety requirements and user preferences, and ensures coherent decision-making across the distributed architecture.
2Reliability
If the first large language model is trained with extensive training data, then the driving performance is improved, but the training data integrity may be compromised by user inputs
Solution Approach 1:
The patent separates the training data management for the first large language model from user input processing. The first model is trained exclusively on curated driving-related data, while user inputs are processed by the second model. This segmentation protects the integrity of the training data while still allowing the system to adapt to user preferences through the second model.
Solution Approach 2:
The intermediary module acts as a barrier that prevents user inputs from directly contaminating the training data of the first large language model. It filters and routes information appropriately, ensuring that only validated, relevant information reaches the first model for potential incorporation.
3Speed
If user inputs are immediately processed by the large language model, then the response speed is fast, but the accuracy of understanding user desires may be insufficient
Solution Approach 1:
The patent divides the input processing into two stages: the second large language model quickly captures and interprets user intent from voice or text inputs, while the first large language model provides contextual validation and refinement. This segmented approach enables fast initial response while maintaining high accuracy through collaborative processing.
Solution Approach 2:
The system implements a feedback loop where the second model's initial interpretation of user inputs is validated and refined by the first model. This feedback mechanism ensures that user desires are accurately understood before being translated into driving actions, improving both precision and overall system reliability.
Data Source
AI summary
For at least partially autonomously driving the motor vehicle, a (second) large language model is interposed to a used artificial intelligence (formed as a first large language model). In this manner, it is possible, to pose queries to the user, who can actively change driving of the motor vehicle via the second large language model. Herein, it can be provided that the artificial intelligence (the first large language model) is exactly not retrained.