Current guard models for language models predict a single 'safe' or 'unsafe' token from the input, concentrating supervision on that one target and leading to overconfidence, shortcut learning, and sensitivity to where evidence appears. LLaDA-Guard instead uses a masked to reconstruct the moderated text under each possible safety label, then classifies based on which label provides a better explanation of the full content.
This generative approach spreads supervision across every token in the moderated region. Evaluated on seven held-out safety benchmarks, LLaDA-Guard achieves better average rank than discriminative baselines trained on stronger backbones, with improved (ECE 0.0875 vs. 0.1384 for Qwen3Guard), reduced over-defense on benign prompts with unsafe-looking cues, and less prompt leakage when moderating responses. The method also produces token-level risk localization and can rewrite unsafe prompts into safe equivalents at a 60.7% conversion rate without additional training.