US20260330291
2026-10-22
Physics
G06N20/00
The patent application discusses a method for aligning machine learning models with human preferences using a technique called optimal transport. This approach responds to a prompt by generating a current response and comparing it with a reference response and preference data. The method involves determining positive and negative reward distributions based on these responses and issuing optimization instructions to align the model with the preference data.
The invention addresses the growing need for control over AI systems, particularly large language models (LLMs), which are capable of generating human-like text. These models are trained on extensive datasets and used in various applications, such as chatbots. Aligning these models with human values and goals is crucial to ensure they are helpful, safe, and reliable, reducing the risk of generating harmful or biased content.
The method involves executing a machine learning model to output a response to a given prompt. A reward component determines positive and negative reward distributions by comparing the current response with a reference response and preference data. An optimizing instruction is issued to align the model according to an optimal transport solution, ensuring that the model's outputs are consistent with human preferences.
This approach is particularly relevant for industries such as healthcare, law, and finance, where the consequences of misleading AI outputs can be severe. By aligning models with human preferences, the method aims to improve the safety and reliability of AI systems. The use of optimal transport allows for efficient processing and resource management, enhancing the model's performance without requiring extensive computation.