Certainty-Flow Tuning for Calibrated Instruction Refusal Behavior in Large Language Models
INTERNATIONAL JOURNAL OF RESEARCH & SOCIAL SCIENCES · 11.09.2026 · 6 (2) · 141-151
DOI: 10.66238/ijrss.v6i2.134 ↗Kurz gefasst
Fachartikel in der Zeitschrift INTERNATIONAL JOURNAL OF RESEARCH & SOCIAL SCIENCES, online erschienen am 11.09.2026. Verfasst von Daphne Hung und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
The rapid advancement of large language models has brought unprecedented capabilities alongside significant challenges in safety alignment. Current alignment techniques frequently result in miscalibrated refusal behaviors, where models either comply with malicious instructions or overly refuse benign queries due to superficial structural similarities to unsafe prompts. This paper introduces certainty-flow tuning, a novel framework designed to achieve calibrated instruction refusal behavior by aligning model responses with their internal confidence regarding prompt safety. The certainty-flow mechanism evaluates the layer-wise propagation of uncertainty during the forward pass, adjusting the gradient updates to penalize unjustified confidence in refusal for ambiguous or benign instructions. Through extensive theoretical analysis and empirical evaluation, this research demonstrates that certainty-flow tuning significantly reduces false refusal rates while maintaining robust defenses against adversarial attacks. The proposed methodology captures the epistemic uncertainty inherent in safety evaluations, translating it into a continuous flow metric that modulates the refusal token generation probability. Our findings suggest that integrating certainty-flow metrics into the fine-tuning pipeline offers a highly effective paradigm for balancing helpfulness and harmlessness, ensuring that language models refuse instructions only when internal safety boundary confidence exceeds dynamically calibrated thresholds.
Zitieren
APA
Hung, D., & Wan, J. (2026). Certainty-Flow Tuning for Calibrated Instruction Refusal Behavior in Large Language Models. INTERNATIONAL JOURNAL OF RESEARCH & SOCIAL SCIENCES, 6(2), 141-151. https://doi.org/10.66238/ijrss.v6i2.134
Vancouver
Hung D, Wan J. Certainty-Flow Tuning for Calibrated Instruction Refusal Behavior in Large Language Models. INTERNATIONAL JOURNAL OF RESEARCH & SOCIAL SCIENCES. 2026;6(2):141-151. doi:10.66238/ijrss.v6i2.134
BibTeX
@article{Hung2026,
author = {Hung, Daphne and Wan, Joyce},
title = {Certainty-Flow Tuning for Calibrated Instruction Refusal Behavior in Large Language Models},
journal = {INTERNATIONAL JOURNAL OF RESEARCH & SOCIAL SCIENCES},
year = {2026},
volume = {6},
number = {2},
pages = {141--151},
doi = {10.66238/ijrss.v6i2.134}
}
EUVIMED