LeanGuard: Fast, Lightweight AI Moderation Without Chain-of-Thought Reasoning.

Current safety guardrail methods for screening prompts and responses often rely on generating a chain-of-thought (CoT) before making a decision, based on the belief that step-by-step reasoning improves accuracy. However, this approach makes guardrails heavy and slow, which is problematic for on-device deployment, such as on embodied robots. This research investigates whether reasoning is truly essential for safety guardrails. By training a lightweight bidirectional encoder and a reasoning guard on the same corpus and then removing only the reasoning component, the study found that CoT does not improve moderation accuracy. The resulting guard, named LeanGuard, is a 395M label-only encoder that achieves an average F1 score of 82.90 across public benchmarks. LeanGuard matches the performance of much larger, decoder-based reasoning guards while requiring only a single forward pass over inputs, leading to approximately a 100x reduction in inference compute. Furthermore, LeanGuard proves more robust under training-label noise and maintains higher recall at strict false-positive rates. This suggests that current guardrail benchmarks might not adequately reward reasoning, and the necessity of CoT for moderation remains unproven. The source code and models are publicly available.