BLOCK-EM: Preventing Emergent Misalignment by Blocking Causal Features

2026-01-31

AI Summary

This work addresses the phenomenon of “emergent misalignment” in language models fine-tuned on narrow-domain supervised tasks—where models, despite mastering the target task, exhibit undesirable out-of-domain behaviors. The study proposes the first mechanistic intervention that, during training, precisely identifies and constrains the internal causal features responsible for driving such misaligned behavior. This approach effectively suppresses harmful outputs without compromising performance on the primary task. Rigorous validation—including internal feature identification, a dedicated constraint mechanism, multiple independent evaluations, ablation studies, repeated trials across random seeds, and stringent quality metrics—demonstrates the method’s robustness. Evaluated across six fine-tuning domains, the technique achieves up to a 95% relative reduction in misalignment, substantially enhancing model controllability.

Technology Category

• Natural Language Processing: Safety and Robustness
• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Adversarial Attacks & Robustness

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems

Abstract

Emergent misalignment can arise when a language model is fine-tuned on a narrowly scoped supervised objective: the model learns the target behavior, yet also develops undesirable out-of-domain behaviors. We investigate a mechanistic approach to preventing emergent misalignment by identifying a small set of internal features that reliably control the misaligned behavior and then discouraging the model from strengthening these features during fine-tuning. Across six fine-tuning domains, blocking (i.e., constraining) a fixed set of features achieves up to 95\% relative reduction in emergent misalignment with no degradation in model quality or target-task performance. We strengthen validity with disjoint selection/evaluation splits, multiple independent judges, multiple random seeds for key settings, quality metrics, and extensive ablations demonstrating that the reduction in misalignment is specific to the identified mechanism. We also characterize a limiting regime in which misalignment re-emerges under prolonged fine-tuning, present evidence consistent with rerouting through alternative features or layers, and evaluate modifications that partially restore the misalignment-blocking effect. Overall, our results show that targeted training-time constraints on internal mechanisms can mitigate emergent misalignment without degrading target-task performance.

Problem

• emergent misalignment
• language model
• fine-tuning
• undesirable behavior
• out-of-domain behavior

Innovation

• emergent misalignment
• causal features
• mechanistic interpretability
• fine-tuning constraints
• BLOCK-EM