Research on Bias Drift and Evaluation of Content Generated by Large Language Models in the Chinese Context
International Journal of Computer Science and Information Technology · 30.10.2026 · 8 (7) · 116-130
DOI: 10.62051/ijcsit.v8n7.12 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Computer Science and Information Technology, online erschienen am 30.10.2026. Verfasst von Tailin Liu. Die Arbeit nennt 18 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).
Abstract (Originalsprache)
Bias in content generated by large language models is not static; it may shift as models evolve, affecting the fairness and reliability of their outputs. Evaluating the presence and patterns of bias drift is therefore important for supporting model safety and fairness. This study examined models from different versions and models with different parameter sizes within the same model family. Using a Chinese-language bias evaluation dataset and metrics designed to assess bias drift, it analyzed changes in bias in model-generated content across versions and parameter sizes. The results showed that both model-version updates and changes in parameter count led to bias drift in generated content, and that the direction of drift varied. A single evaluation or assessments conducted only at fixed intervals may not detect bias drift promptly as large language models evolve. Version updates and increases in parameter count should therefore not, by themselves, be taken as evidence of improved model safety; ongoing bias evaluation and verification of alignment effectiveness are needed. Value alignment is not a one-time process, and its effectiveness cannot be assumed to persist over time. Alignment effectiveness should be assessed throughout model-version updates and changes in parameter count, with verification serving as a necessary checkpoint after a version update or an increase in parameter count. This study provides empirical evidence for evaluating and analyzing bias drift in content generated by large language models and may inform research on model alignment and the development of regulatory frameworks.
Zitieren
APA
Liu, T. (2026). Research on Bias Drift and Evaluation of Content Generated by Large Language Models in the Chinese Context. International Journal of Computer Science and Information Technology, 8(7), 116-130. https://doi.org/10.62051/ijcsit.v8n7.12
Vancouver
Liu T. Research on Bias Drift and Evaluation of Content Generated by Large Language Models in the Chinese Context. International Journal of Computer Science and Information Technology. 2026;8(7):116-130. doi:10.62051/ijcsit.v8n7.12
BibTeX
@article{Liu2026,
author = {Liu, Tailin},
title = {Research on Bias Drift and Evaluation of Content Generated by Large Language Models in the Chinese Context},
journal = {International Journal of Computer Science and Information Technology},
year = {2026},
volume = {8},
number = {7},
pages = {116--130},
doi = {10.62051/ijcsit.v8n7.12}
}
EUVIMED