Refusal-Aware Instruction Tuning for Hallucination Mitigation near Uncertain Knowledge Boundaries
International Journal of Information and Communication Technology Trends · 28.07.2026 · 6 (2) · 91-99
DOI: 10.66525/ijictt111 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Information and Communication Technology Trends, online erschienen am 28.07.2026. Verfasst von Zoé Mathieu und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
The rapid deployment of large language models has exposed significant vulnerabilities in their reliability, particularly concerning their propensity to hallucinate when confronted with queries that exceed their internalized knowledge capacities. This paper investigates the critical challenge of mitigating hallucinations near uncertain knowledge boundaries through a novel paradigm termed refusal-aware instruction tuning. Conventional instruction tuning often inadvertently encourages models to generate responses even when uncertain, prioritizing fluency and helpfulness over factual accuracy. To counteract this phenomenon, we propose a systematic framework that integrates uncertainty quantification directly into the instruction tuning pipeline, teaching models to explicitly refuse to answer when a query falls outside their reliable knowledge distribution. By formalizing the concept of uncertain knowledge boundaries, we develop a specialized dataset curation strategy that identifies low-confidence regions and pairs them with appropriate refusal responses. Our methodology is subsequently evaluated through extensive empirical analysis, demonstrating that refusal-aware instruction tuning substantially reduces hallucination rates without compromising the utility of the model on well-known facts. The findings suggest that aligning models with their own epistemic limitations is a fundamental requirement for deploying trustworthy artificial intelligence systems in high-stakes domains.
Zitieren
APA
Mathieu, Z., & Roy, A. T. (2026). Refusal-Aware Instruction Tuning for Hallucination Mitigation near Uncertain Knowledge Boundaries. International Journal of Information and Communication Technology Trends, 6(2), 91-99. https://doi.org/10.66525/ijictt111
Vancouver
Mathieu Z, Roy AT. Refusal-Aware Instruction Tuning for Hallucination Mitigation near Uncertain Knowledge Boundaries. International Journal of Information and Communication Technology Trends. 2026;6(2):91-99. doi:10.66525/ijictt111
BibTeX
@article{Mathieu2026,
author = {Mathieu, Zoé and Roy, Adrien T.},
title = {Refusal-Aware Instruction Tuning for Hallucination Mitigation near Uncertain Knowledge Boundaries},
journal = {International Journal of Information and Communication Technology Trends},
year = {2026},
volume = {6},
number = {2},
pages = {91--99},
doi = {10.66525/ijictt111}
}
EUVIMED