DT-Guard: Intent-Driven Reasoning-Active Training for Reasoning-Free LLM Safety Guardrail

2026-07-07

AI Summary

This work addresses the trade-off between efficiency and nuanced risk detection in safety guardrails for large language models, where lightweight classifiers often fail to capture implicit malicious intent, while reasoning-based approaches incur high latency. To overcome this limitation, the authors propose DT-Guard, a novel framework that adopts a “reasoning-at-training, inference-free-at-deployment” paradigm. DT-Guard employs a structured three-stage decision pipeline—intent recognition, categorization, and safety assessment—during training, internalizing complex reasoning into the model parameters. It further incorporates Rollout-Guided Progressive Hard-Case Optimization to specifically refine performance on challenging examples. Evaluated on both prompt-side and response-side safety benchmarks, DT-Guard achieves average F1 scores of 0.886 and 0.870, respectively, with an overall score of 0.878—surpassing an 8B-parameter baseline despite using only a 4B-parameter model—demonstrating both high efficiency and robust safety judgment.

Abstract

Large language models deployed in open-world applications require safety guardrails that are both robust to complex risks and efficient enough for low-latency runtime moderation. Existing guardrails face a practical trade-off between lightweight classification-based models, which are efficient but often struggle with concealed intent, ambiguous semantics, and borderline safety decisions, and reasoning-based guards, which improve judgment quality but introduce additional token generation and inference latency. We present DT-Guard, a content safety guardrail model based on a Reasoning-Active Training, Reasoning-Free Inference paradigm. The key idea is to use reasoning supervision during training while emitting only structured safety labels at inference time. DT-Guard formulates safety judgment as a progressive decision process, Intent - Category - Safety, and constructs an intent-driven dataset with intent labels, risk categories, safety labels, and structured reasoning trajectories. To further improve hard-case robustness, we propose Rollout-Guided Progressive Hard-Case Optimization (RG-PHO), which uses multi-rollout consistency to identify stably mastered, persistently failed, and preference-unstable samples, and applies targeted supervised and preference optimization accordingly. At inference time, DT-Guard directly generates structured labels without explicit reasoning traces, preserving deployment efficiency. Experiments on prompt-side and response-side safety benchmarks show that DT-Guard achieves average F1 scores of 0.886 and 0.870, respectively. With only a 4B backbone, it reaches a dual-side average F1 of 0.878, outperforming strong 8B guardrail baselines. These results demonstrate that reasoning supervision can be effectively internalized into low-latency safety discrimination.

Problem

• safety guardrail
• reasoning-free inference
• concealed intent
• low-latency moderation
• structured safety labels

Innovation

• Reasoning-Active Training
• Reasoning-Free Inference
• Intent-Driven Safety Guardrail
• Hard-Case Optimization
• Structured Safety Labeling