Robust and Efficient Guardrails with Latent Reasoning

The researchers proposed the COLAGUARD guard model, which aims to transfer multi-step safe reasoning to continuous latent spaces through a phased training curriculum and directly hide state propagation during reasoning. The model was evaluated using ten prompt and response adjustment settings covering eight safety benchmarks, and results showed that its macro F1 score increased by 8.24 points compared to Llama Guard 3, reaching the level of the explicit reasoning baseline GuardReasoner. Additionally, COLAGUARD increased the reasoning speed by 12.9 times and reduced token usage by 22.4 times. The findings indicate that latent reasoning provides a practical alternative for deployable safety guards, capable of jointly enhancing safety robustness and reasoning efficiency.