IHDec: Divergence-Steered Contrastive Decoding for Securing Multi-Turn Instruction Hierarchies

Large language models (LLMs) have revolutionized natural language processing, enabling complex multi-source, multi-turn interactions. However, maintaining strict instruction hierarchies—where higher-priority directives override subordinate ones—remains a significant challenge, especially in multi-turn dialogues. Existing solutions often rely on costly fine-tuning or prompt engineering, which are limited in scalability and robustness. This paper introduces IHDec, a novel inference-time framework that leverages Jensen-Shannon Divergence (JSD) to quantify role influence during generation. By monitoring divergence scores, IHDec detects potential hierarchy violations at token level and dynamically adjusts logits through contrastive decoding, suppressing subordinate role interference. Extensive evaluations on the IHEval benchmark demonstrate that IHDec outperforms state-of-the-art training-based methods, achieving an 11.98 percentage point improvement in instruction adherence in conflict scenarios. It also enhances safety against adversarial prompts, with accuracy rising to 54.3%. The approach scales effectively with larger models, such as Qwen3-32B, where performance improves monotonically with size, reaching a 27.8-point gain. This method’s model-agnostic, training-free nature makes it highly practical for deployment across diverse architectures, offering a significant step toward safer, more reliable AI systems. Future work will explore reinforcement learning integration and multi-modal extensions, aiming to further improve robustness and efficiency in complex real-world applications.