New MOOD benchmark boosts LLM safety by detecting unknown failures

Researchers from the paper titled "Benchmarking and Improving Monitors for Out-Of-Distribution Alignment Failure in LLMs" introduce MOOD (Misalignment Out Of Distribution), a dedicated benchmark to systematically evaluate whether LLM monitoring pipelines can detect safety failures that occur when models encounter unusual prompts or response patterns—i.e., out-of-distribution (OOD) scenarios. Standard safety classifiers (guard models) are trained on vast datasets but often fail to generalize to novel failure modes. The MOOD benchmark sidesteps this by providing a restricted training set to train custom monitors alongside seven diverse test sets containing alignment failures that are genuinely outside the training distribution.

Using MOOD, the researchers found that off-the-shelf guard models significantly underperform on OOD failures—recall as low as 39%. To address this, they propose combining guard models with OOD detectors. Testing four detector types, they show that pairing a guard model with Mahalanobis distance and perplexity-based OOD detectors lifts recall to 45%—a 15% relative improvement. Remarkably, this combination outperforms using a guard model alone that has 20 times more parameters. The work establishes positive scaling trends and suggests that OOD detection should be a standard component of LLM safety monitoring.