Patents

US 12724752 B2: System and method for generating thoughts with large language models using codewords

Granted
2026-09-01
Filed
2025-11-21 application 19397339
Type
Utility · 20 claims
Elsewhere
Google Patents ↗ · USPTO patent document ↗ · USPTO Patent Center record ↗

Abstract

This invention presents an optimized approach for training and operating Large Language Models (LLMs) using codewords. By converting traditional token-based LLMs to codeword-based systems, the method achieves significant efficiency gains. The process involves tokenizing training data and assigning codewords to tokens. LLMs are then trained and operated using these compact codewords instead of conventional tokens. During operation, prompts are converted to codewords, processed by the LLM, and the outputs are converted back to text. This approach reduces the overall cost of training and operating LLMs by approximately, offering a more efficient solution for large-scale language processing tasks.

Source: PTBLXML, file of 2026-09-01 (ipgb20260901.xml).

Assignees

AssigneeTypePlaceAlso on furp.orgSource
ATOMBEAM TECHNOLOGIES INC. 213 patentsUS company or organizationMoraga, CAATOMBEAM TECHNOLOGIES INC. State business registration EXACT high confidence: identical full name and state
AtomBeam Technologies Inc. Form D issuer EXACT high confidence: identical full name and state
PTBLXML, file of 2026-09-01

Inventors

InventorPlacePatentsPerson pageSource
Brian GalvinSilverdale, WA128PTBLXML, file of 2026-09-01
Alan McCordForney, TX36PTBLXML, file of 2026-09-01

CPC categories

SubclassCategoryFirst groupSource
G06FElectric digital data processingG06F16/212PTBLXML, file of 2026-09-01

Cooperative Patent Classification, as currently assigned (one row per subclass).

Assignments and security interests

No recorded assignments or security interests on file for this patent.

USPTO bulk data: PatentsView granted patents, disambiguated (PVGPATDIS) for patents, assignees, inventors, places and CPC categories; the USPTO Patent Assignment Dataset (ECORSEXC) for assignments and security interests. Patents filed in the last 20 years. No claims or full text. Links to other records are by name (and state): EXACT means the identical full name and state; FUZZY means a looser name match, with the confidence shown.