Abstract
Conversational large language models (LLMs) often have limited capabilities when functioning as a real-time interpreter. These models can fail to follow instructions consistently, resulting in responses in the incorrect language or attempts to answer a user's question instead of providing a direct translation. This issue stems from a lack of training data that reflects the typical turn-by-turn, interleaved conversation flow between two speakers of different languages. Some implementations are directed to transitioning a generative model-powered assistant into a translation mode in which the assistant facilitates structured conversational turns between two users speaking different languages, with the generative model acting as the interpreter that translates between two languages used by the two users. Additionally or alternatively, some implementations are directed to creating training data that adapts a generative model to effectively act as a translator.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Hard, Andrew; Zhang, Shujian; Barua, Aditya; Lambert, John; and Zhou, Wenxuan, "Transitioning a Generative Model-Powered Assistant into Translation Mode", Technical Disclosure Commons, (September 18, 2025)
https://www.tdcommons.org/dpubs_series/8604