LPG: Balancing Efficiency and Policy Reasoning in Latent Policy Guardrails

In three linesLPG (Latent Policy Guardrail) is a safety framework for LLMs that adapts security policies at inference time without retraining. It compresses reasoning into 10 latent tokens, achieves 84.5% accuracy and 77.9% F1 on benchmarks, while running 11× faster than Qwen3-4B-Thinking.

Summary generated by Claude — human-verified