Breakthrough Patent Application for Improved Language Models

Strategic Insights -

Patent inventors Egerton, Lauren Elizabeth and colleagues at THIA ST Co. have submitted a groundbreaking patent application for advanced language models that can process text inputs more efficiently, while reducing computational burdens and artifact generation. The novel microservice architecture utilises small to mid-size trained machine learning tools, offering lower computational effort and bespoke deployment options. This technology has significant implications for applications in finance, legal scholarship, programming, and chatbots.

Key Takeaways:

  • The patent application aims to address limitations of current large language models (LLMs), including computational burden and undesirable artifacts.
  • The inventors propose a microservice architecture that incorporates small to mid-size trained machine learning tools for improved efficiency and customization.
  • The technology can be deployed on workstation or single-node computer systems with a single GPU, reducing the need for expensive clusters.
  • Multiple pretraining stages can be performed on an expansion LLM to expand the reach of client inputs and provide better responses.
  • Retrieval augmented generation (RAG) and data producer modules can be used to enhance the pool of documents and knowledge graphs for addressing client inputs.
  • The system can retain histories for multiple client entities in respective long-term memories, improving response accuracy and relevance.
  • The patent application explores the application of this technology in various domains, including finance, law, and programming.

Statistics:

  • The patent application was filed on June 24, 2025, and made available online on October 16, 2025.
  • The novel microservice architecture utilises small to mid-size trained machine learning tools, reducing computational effort.
  • The system can be deployed on workstation or single-node computer systems, reducing the need for expensive clusters.
  • The patent application cites the benefits of using an expansion LLM with multiple pretraining stages for improved response accuracy and relevance.

Sources:

  • Egerton, Lauren Elizabeth; Kelly, Brendan Michael, Kelsey, Elaine; Nasir, Sazzad Mahmud, Robson, Elliot Nicholas, Robson, Robert Oscar, Ward, Spencer Thomas, Yarbro, Jeffrey Thomas. Copilot Implementation: Data Retrieval Over Application Programming Interface (Api). U.S. Patent Application Number 20250321977, filed June 24, 2025, and posted October 16, 2025.
  • Patent URL (for desktop use only): https://ppubs.uspto.gov/pubwebapp/external.html?q=(20250321977)&db=US-PGPUB&type=ids
  • Keywords: Business, THIA ST Co, Library Science, Data Acquisition, Information Technology.