Tag
CoLaGuard is a new guardrail model that transfers multi-step safety reasoning into a continuous latent space, achieving 12.9x speedup and 22.4x token reduction compared to explicit reasoning baselines while matching macro-F1 performance on ten safety benchmarks.