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

Source

Source name: arXiv
Source kind: paper
Authors / organization: He Liu, Changtao Miao, Xinjie Yang, Tianle Song, Yin Wu, Junchi Chen, Bintao He, Xinyuan Zhang, Bo Zhang, Shi Yan, Wei Lu, Wei Wang, Danyang Xu, Jiansheng Cai, Zhe Li
Venue: arXiv: cs.AI
Published: 2026-07-07
Collected: 2026-07-08
URL: http://arxiv.org/abs/2607.06326v1
PDF: https://arxiv.org/pdf/2607.06326v1
DOI: UNKNOWN

Plain-Language Brief

Researchers present DT-Guard, a content safety guardrail model for large language models in open-world applications. The model uses a Reasoning-Active Training, Reasoning-Free Inference paradigm and aims to improve robustness to complex risks while maintaining efficiency. It formulates safety judgment as a progressive decision process and uses Rollout-Guided Progressive Hard-Case Optimization (RG-PHO) for hard-case robustness. DT-Guard outperforms strong 8B guardrail baselines on prompt-side and response-side safety benchmarks.

What the Document Says

DT-Guard is a content safety guardrail model
The model uses a Reasoning-Active Training, Reasoning-Free Inference paradigm
Safety judgment is formulated as a progressive decision process: Intent - Category - Safety
An intent-driven dataset with intent labels, risk categories, safety labels, and structured reasoning trajectories is constructed
Rollout-Guided Progressive Hard-Case Optimization (RG-PHO) is proposed for hard-case robustness
DT-Guard outperforms strong 8B guardrail baselines on prompt-side and response-side safety benchmarks
The average F1 scores are 0.886 for prompt-side and 0.870 for response-side, and 0.878 for dual-side with a 4B backbone

Key Concepts Mentioned

DT-Guard
Reasoning-Active Training
Reasoning-Free Inference
Safety guardrail model
Rollout-Guided Progressive Hard-Case Optimization (RG-PHO)
F1 scores

Claims Made by the Document

DT-Guard improves robustness to complex risks while maintaining efficiency
DT-Guard outperforms strong 8B guardrail baselines on prompt-side and response-side safety benchmarks

Evidence Used

The text describes DT-Guard's paradigm, formulation of safety judgment, dataset construction, RG-PHO, and experimental results

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

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.