DT-Guard Enhances LLM Safety with Reasoning-Active Training.

Deploying large language models in open environments necessitates robust and efficient safety guardrails. Existing solutions often face a trade-off: lightweight classification models are fast but struggle with nuanced risks, while reasoning-based guards offer better judgment but introduce latency. DT-Guard, a novel content safety guardrail model, resolves this by employing a "Reasoning-Active Training, Reasoning-Free Inference" approach. During training, it leverages reasoning supervision, formulating safety judgment as a progressive decision process (Intent - Category - Safety) and constructing a dataset with detailed reasoning trajectories. To further boost robustness, DT-Guard uses Rollout-Guided Progressive Hard-Case Optimization (RG-PHO) to identify and target difficult samples for optimization. Crucially, at inference time, DT-Guard directly outputs structured safety labels without generating explicit reasoning traces, maintaining deployment efficiency. Benchmarks show DT-Guard, even with a smaller backbone, outperforms larger baselines in F1 scores for both prompt and response-side safety.