Abstract
Proposed herein is a system that enables Large Language Models (LLMs) to automatically generate linguistically appropriate hedging language that accurately reflects the reliability of underlying information sources. By extracting multi-dimensional epistemological metadata (source authority, temporal relevance, evidential support) from knowledge base documents and integrating this metadata directly into LLM generation processes through augmented attention mechanisms, the system produces responses where confidence expressions match actual source reliability. The approach addresses the "confidently incorrect" problem in high-stakes applications (medical, legal, financial) by ensuring LLMs communicate uncertainty appropriately, using domain-specific hedging conventions and continuous feedback-driven refinement.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Steimel, Kenneth; Sardana, Aman; and Dozier, Scott, "SYSTEM FOR EPISTEMOLOGICALLY-INFORMED TEXT GENERATION WITH LANGUAGE MODELS", Technical Disclosure Commons, (January 12, 2026)
https://www.tdcommons.org/dpubs_series/9172