Robust and Efficient Guardrails with Latent Reasoning

In three linesCOLAGUARD, a guardrail model, transfers multi-step safety reasoning into continuous latent space via stage-wise training curriculum. Evaluated on 10 moderation tasks across 8 safety benchmarks, it improves macro-F1 by 8.24 points over Llama Guard 3, matches GuardReasoner performance while delivering 12.9X speedup and 22.4X token reduction.

Summary generated by Claude — human-verified