Do Safety Guardrails Need to Reason? LeanGuard: A Fast and Light Approach for Robust Moderation

This paper investigates whether the prevailing trend of using chain-of-thought (CoT) reasoning in safety guardrails is actually necessary for accurate content moderation. While many modern guardrails force models to generate a reasoning trace before issuing a verdict, this process is computationally expensive and slow. The author performs a controlled, same-base comparison by training models with and without CoT on identical datasets and architectures. By isolating the effect of the reasoning step, the study evaluates whether CoT provides a genuine accuracy gain or if it is merely a post-hoc justification for a decision already reached by the model.

The study demonstrates that CoT does not improve moderation accuracy in a controlled setting. On the same model backbone, removing the reasoning step often results in identical or even superior performance. The author introduces LeanGuard, a 395M-parameter bidirectional encoder that performs single-pass classification. LeanGuard achieves an average F1 score of 82.90, matching or exceeding much larger reasoning-based decoders while reducing inference compute by approximately 100x. Furthermore, the author finds that reasoning-based guards often exhibit polarized confidence, which can hinder their performance at strict false-positive rate thresholds, whereas the discriminative encoder remains more robust to training-label noise.

These findings challenge the current consensus that reasoning is essential for safety moderation. By showing that a lightweight, single-pass classifier is sufficient, the paper provides a practical, high-performance alternative for on-device deployment, such as in robotics or edge computing, where latency and compute resources are constrained. The results suggest that the field may be over-investing in complex generative architectures for tasks that are fundamentally bounded-label classification problems.