BLOCK-EM: Preventing Emergent Misalignment via Latent Blocking

TL;DR

This paper presents a mechanistic method called BLOCK-EM that significantly reduces emergent misalignment in language models by constraining internal features during fine-tuning, without harming performance.

Contribution

It introduces a targeted feature-blocking approach that prevents undesirable behaviors emerging during model fine-tuning, validated across multiple domains and conditions.

Findings

Up to 95% reduction in emergent misalignment.

No degradation in model quality or target-task performance.

Misalignment re-emerges under prolonged fine-tuning, indicating limits of the method.

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.