LPG: Balancing efficiency and policy reasoning in latent policy guardrails

The LPG (Latent Policy Guardrail) framework compresses internal deliberation under dynamic safety policies into latent states, and at inference outputs only a short verdict anchored to the violated clause. On the policy guardrail benchmark, LPG-4B achieves 84.5% average safety accuracy and 77.9% F1 with 10 latent tokens, surpassing the strongest dynamic baseline. Under the single-sample evaluation setting, it is approximately 11 times faster than Qwen3-4B-Thinking.