Benchmarking and Improving Monitors for Out-Of-Distribution Alignment Failure in LLMs

Dylan Feng, Pragya Srivastava, Cassidy Laidlaw

arXiv:2605.21602 · 2026-05-23 공개 · arXiv · PDF

llm-safety model-scaling mahalanobis-distance ood-detection alignment-failure guard-models llm-monitoring safety-classifiers

Abstract

Many safety and alignment failures of large language models (LLMs) occur due to out-of-distribution (OOD) situations: unusual prompt or response patterns that are unforeseen by model developers. We systematically study whether LLM monitoring pipelines can detect these OOD alignment failures by introducing a benchmark called Misalignment Out Of Distribution (MOOD). It is difficult to find failures that are truly OOD for off-the-shelf models trained on vast safety datasets. We sidestep this by including a restricted training set in MOOD that we use to train our own monitors, as well as seven test sets with diverse alignment failures that are outside the training distribution. Using MOOD, we find that guard models (safety classifiers) often fail to generalize OOD. To fix this, we propose combining guard models with OOD detectors. We test four types of OOD detectors and find that a combination of a guard model with Mahalanobis distance and perplexity-based OOD detectors can improve recall from 39% to 45%. We also establish positive scaling trends across model scales for monitors that combine a guard model and OOD detector; we find that incorporating OOD detection into monitoring achieves a higher recall gain than using a guard model with 20 times more parameters. Our work suggests that OOD detection should be a crucial component of LLM monitoring and provides a foundation for further work on this important problem.

한국어 요약

📋 한 줄 요약

**[LLM Safety / OOD Alignment Monitoring]** MOOD 벤치마크로 guard model의 OOD 일반화 실패 진단, Mahalanobis distance·perplexity 기반 OOD detector 결합 시 recall 39%→45%, 20× 큰 guard model보다 OOD detector 결합이 더 큰 recall 향상.

🎯 핵심 기여도

💡 핵심 아이디어

LLM monitoring은 단일 guard model로는 OOD 실패에 generalize되지 않으며, guard model과 OOD detector(Mahalanobis·perplexity)의 결합이 모델 scaling보다 효율적인 recall 향상 경로다 — OOD detection이 LLM monitoring의 crucial component로 자리잡아야 한다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: LLM safety monitoring에 OOD detection을 핵심 component로 정립, MOOD로 후속 연구 기반 제공, scaling보다 효율적인 monitoring 향상 경로 입증. **한계**: 자체 학습 monitor의 일반화는 MOOD 외 distribution에서 검증 필요, recall 39→45는 절대값으로는 여전히 개선 여지, 4 detector type 외 더 정교한 OOD 방법 미탐색.

🚀 실용적 활용