HiRoute Enhances LLM Safety Alignment with Hierarchical Prompt Tuning.

Large language models (LLMs) frequently face challenges with harmful requests and jailbreak attempts, and current safety alignment methods often struggle to balance robust safety with helpful, non-over-refusing responses. Existing prompt tuning techniques typically use static, global prompts or pre-selected modules, which can be inflexible. A new framework, HiRoute, addresses these issues by introducing an input-adaptive hierarchical prompt-tuning approach. This method intelligently separates the general safety control, which is category-agnostic, from specific response guidance tailored to particular risks. It employs a lightweight hierarchical router to detect harmful intent and predict risk scores, then uses a combination of a shared coarse-grained prompt and risk-specific prompt experts. This dynamic system ensures that benign inputs bypass the safety branch, while risky inputs are processed with a weighted mixture of expert prompts. Experimental results across various instruction-tuned models demonstrate that HiRoute significantly improves safety rates on benchmarks, preserves helpfulness, reduces unnecessary refusals, and maintains strong performance on general tasks.