According to arxiv.org, researchers have developed COLAGUARD, a new guardrail model that transfers multi-step safety reasoning into a continuous latent space, addressing the challenge of high latency in existing safety systems. The model improves macro-F1 scores by 8.24 points over Llama Guard 3 while delivering a 12.9X speedup and 22.4X reduction in token usage compared to explicit reasoning baselines.
In parallel work, arxiv.org reports the release of Opir, a family of encoder-based guardrail models built on the GLiClass architecture for real-time safety filtering. Opir includes variants with fewer than 100M parameters for binary safe/unsafe categorization and supports multi-task classification across toxicity, jailbreak attempts, and unsafe responses. According to the paper, the models are trained on a three-level taxonomy containing 996 categories and perform competitively against eight contemporary guardrail systems.
Additionally, arxiv.org describes DenseSteer, a training-free inference-time steering framework that enhances small language models (≤3B parameters) on mathematical reasoning tasks by modulating internal representations toward “dense reasoning patterns”—fewer steps with higher information density per step.
For diffusion language models specifically, arxiv.org introduces DLM-SWAI, a training-free steering method that biases token distribution during iterative denoising using pre-computed token-level style scores for style and safety control.