Abstract
Traditional Input Method Editors (IMEs) that provide next phrase prediction to ease data entry have limited context and lack deep semantic understanding. The prediction accuracy of such IMEs is suboptimal for long-tail phrases such as domain-specific vocabulary, acronyms, slang, etc. or in contexts where the text entry has a long semantic context. This document describes techniques that integrate a Large Language Model (LLM) into the text decoding and prediction path of an IME to reduce keystroke overhead and improve output quality by providing advanced state-dependent predictions. An asynchronous prediction pipeline is implemented where a client application (e.g., that operates as an add-to a standard IME) extracts context for a current token, sends a request to an LLM (e.g., on a remote server or the local device) to obtain predictions, and displays the predictions asynchronously to the user. By decoupling context extraction and prediction requests from a main UI thread of the standard IME, near-real time LLM-powered text predictions can be provided. Predictions from the standard IME continue to be provided and serve as a fallback.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Jiang, Liyang; Su, Heng; and Feng, Gang, "Large Language Model (LLM) Powered Input Method Editor (IME)", Technical Disclosure Commons, (September 02, 2026)
https://www.tdcommons.org/dpubs_series/11563