HeteroGuard: Constructing Safety Boundaries for Large Language Models via Heterogeneous Weak-Model Collaboration
Security and Safety · 30.07.2026
DOI: 10.1051/sands/2026018 ↗Kurz gefasst
Fachartikel in der Zeitschrift Security and Safety, online erschienen am 30.07.2026. Verfasst von Ping Chen und 6 weiteren Autor:innen. Gefördert von: National Key Research and Development Program of China.
Abstract (Originalsprache)
Large language models (LLMs) face safety risks in deployment, but existing evaluation methods rely on single powerful judges and treat safety as binary classification. This paper proposes HeteroGuard, a dynamic heterogeneous redundancy (DHR)-inspired framework that constructs a safety boundary around LLM outputs through heterogeneous weak-model collaboration. HeteroGuard deploys three diverse weak judges and aggregates their decisions via majority voting to define a boundary that organizes responses into risk-stratified regions. Experiments on multiple target models show that responses outside the boundary are harmful with probability $56\%$–$65\%$, corresponding to lifts of $7\times$–$14\times$ compared with the global harmful rate. Vote-stratified harmful probabilities increase monotonically from $1\%$–$4\%$ for zero votes to over $80\%$ for unanimous votes, confirming that weak-judge consensus indicates true harmfulness. HeteroGuard achieves higher precision and lower false-positive rates than individual weak judges, demonstrating that structural diversity enhances evaluator robustness. This work validates applying endogenous security principles to LLM safety evaluation.
Förderung
- National Key Research and Development Program of China
Zitieren
APA
Chen, P., Cai, Y., Zhang, S., Wang, Y., Shen, X., Shang, Y., & Zou, H. (2026). HeteroGuard: Constructing Safety Boundaries for Large Language Models via Heterogeneous Weak-Model Collaboration. Security and Safety. https://doi.org/10.1051/sands/2026018
Vancouver
Chen P, Cai Y, Zhang S, Wang Y, Shen X, Shang Y, et al. HeteroGuard: Constructing Safety Boundaries for Large Language Models via Heterogeneous Weak-Model Collaboration. Security and Safety. 2026. doi:10.1051/sands/2026018
BibTeX
@article{Chen2026,
author = {Chen, Ping and Cai, Yin and Zhang, Shuting and Wang, Yi and Shen, Xingqiu and Shang, Yuting and Zou, Hong},
title = {HeteroGuard: Constructing Safety Boundaries for Large Language Models via Heterogeneous Weak-Model Collaboration},
journal = {Security and Safety},
year = {2026},
doi = {10.1051/sands/2026018}
}
EUVIMED