BLOCK-EM: Preventing Emergent Misalignment via Latent Blocking
Not on the map. Why?
Abstract (arXiv)
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.
Key facts. Cited by 2 papers in the map (data through 7 Oct 2026). Builds on 18 of its 33 references that are in the map. First posted on arXiv 31 Jan 2026.
Builds on
18 of 33 references.
References · 18
- 2022Locating and Editing Factual Associations in GPTcited by 2,609
- 2022Emergent Abilities of Large Language Modelscited by 2,483
- 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,359
- 2023Steering Llama 2 via Contrastive Activation Additioncited by 1,009
- 2024Refusal in Language Models Is Mediated by a Single Directioncited by 960
- 2022Toy Models of Superpositioncited by 928
- 2023Steering Language Models With Activation Engineeringcited by 824
- 2024Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2cited by 381
- 2023Towards Best Practices of Activation Patching in Language Models: Metrics and Methodscited by 328
- 2025Persona Vectors: Monitoring and Controlling Character Traits in Language Modelscited by 303
- 2025Layer by Layer: Uncovering Hidden Representations in Language Modelscited by 293
- 2024How to use and interpret activation patchingcited by 191
6 more
- 2025Persona Features Control Emergent Misalignmentcited by 76
- 2025Thought Crime: Backdoors and Emergent Misalignment in Reasoning Modelscited by 57
- 2025Steering Out-of-Distribution Generalization with Concept Ablation Fine-Tuningcited by 34
- 2025Inoculation Prompting: Instructing LLMs to misbehave at train-time improves test-time alignmentcited by 32
- 2025Interpretable LLM Guardrails via Sparse Representation Steeringcited by 5
- 2025The Devil in the Details: Emergent Misalignment, Format and Coherence in Open-Weights LLMscited by 4
Built on by
Citing papers per year: 2026: 2.
2 papers.