COLAGUARD: Robust and efficient guardrails via latent reasoning

The researchers propose the guardrail model COLAGUARD, which transfers multi-step safety reasoning into a continuous latent space through a staged training curriculum; at inference time it directly propagates hidden states, eliminating the latency and token overhead of explicitly generating reasoning chains. Across ten prompt and response moderation settings covering eight safety benchmarks, COLAGUARD's macro-F1 is 8.24 points higher than Llama Guard 3 and on par with the explicit-reasoning baseline GuardReasoner, while achieving a 12.9× speedup and a 22.4× reduction in token usage. The authors argue that latent reasoning offers a practical path beyond explicit rationale generation for deployable guardrails, improving both safety robustness and inference efficiency.