Abstract
Interactions with large language models are often processed in isolated sessions to manage context constraints and inference costs. This isolation prevents the consolidation of knowledge across different conversations. As a result, generated responses may include redundant information already familiar to a user or present new concepts without adequate context for understanding. This disclosure describes a method where model outputs are decomposed into discrete knowledge blocks based on previous interactions. During the inference process, relevant blocks are retrieved and categorized as either known or novel. Known blocks are referenced to provide context, while new information is explained through connections to established concepts. The ratio of novelty in a response is adjustable through a user interface. By aligning content with a specific knowledge base, the comprehension and verifiability of responses are improved while computational redundancy is reduced.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Tran, Duc-Hieu and Hartmann, Florian, "Decomposing Large Language Model Generations Based on User Knowledge for Controlled Novelty in Responses", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10108