HeteroGuard: Constructing Safety Boundaries for Large Language Models via Heterogeneous Weak-Model Collaboration

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.