Collaborative Dual-Size Large Language Models with Dual-Stage Deferral Risk Control

ICLR 2026 Conference Submission11736 Authors

Keywords: Risk Control, Model Collaboration, Deferral Mechanism, Computational Efficiency, Safety-Efficiency Trade-off, Dual-size Models

Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities, yet ensuring their safe deployment remains challenging. Existing safety mechanisms, while effective against malicious inputs, often degrade performance on benign queries due to over-conservative strategies. We propose the \textbf{D}ual-size LLM collaborative framework with \textbf{D}ual-stage deferral risk contro\textbf{L} (\textbf{DDL}), which integrates lightweight and heavyweight models with calibrated deferral mechanisms. Our approach formalizes the safety–efficiency trade-off as a constrained optimization problem that jointly considers prediction accuracy, computational cost, and safety risk. We provide theoretical guarantees showing that our mechanism achieves distribution-free risk control while minimizing unnecessary heavyweight computation. Extensive experiments on three datasets demonstrate that DDL effectively balances safety and efficiency, achieving performance and safety metrics comparable to state-of-the-art safety-aligned models while reducing average inference time by more than 65\%.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 11736