Research map / Large language models / Language, safety and interpretability
Emergent Misalignment is Easy, Narrow Misalignment is Hard
Anna Soligo, Edward Turner, Senthooran Rajamanoharan, Neel Nanda
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| mix single-domain misaligned data with aligned data from other domains (ablation) | LoRA finetune on misaligned advice · Emergent misalignment | general and narrow misalignment | No clear effect |
| add KL regularisation on out-of-domain data to the finetuning loss (ablation) | steering vector or LoRA finetune · Bad Medical Advice | general misalignment | Improved |
| train narrow solution instead of general solution (efficiency: loss at equal parameter norm) (ablation) | steering vector · Bad Medical Advice | training loss at equal norm | Hurt |
| train narrow solution instead of general solution (stability to orthogonal noise) (ablation) | steering vector · Bad Medical Advice | loss increase under noise | Hurt |
| remove KL regularisation and continue finetuning narrow solution (ablation) | narrow steering vector · Bad Medical Advice | misalignment generality | Hurt |
| steer with general instead of narrow or random vectors (ablation) | misalignment vectors · FineWeb | KL divergence from chat model | Improved |
| use text datasets instead of insecure code for finetuning | Qwen2.5-14B-Instruct · Emergent misalignment | EM and coherency: 6% EM and 67% coherency -> nearing 40% misalignment with over 99% coherency | Improved |
| use full SFT instead of LoRA (ablation) | Qwen2.5-14B-Instruct · Emergent misalignment | EM (%): 8.92 to 36.34 | No clear effect |
| use full SFT on Gemma instead of Qwen | gemma-3-12b-it · Emergent misalignment | EM (%): 3.82 to 13.78 after 3 epochs | Hurt |
| use rank-1 LoRA instead of rank-32 LoRA (ablation) | Qwen2.5-14B-Instruct · Emergent misalignment | EM (%): 9.25 to 21.5 | No clear effect |
| increase model size (ablation) | Qwen and Llama models · Emergent misalignment | EM and coherency: up to 8% and 9% EM at smallest sizes | Improved |
| increase model size (ablation) | Gemma-3 models · Emergent misalignment | EM | No clear effect |
Closest papers by meaning
- LLM Alignment--Utility Asymmetry under Semantic-Preserving Transformations (NeurIPS 2026 · similarity 0.89)
- Alignment Dynamics in LLM Fine-Tuning (NeurIPS 2026 · similarity 0.88)
- Model Spec Midtraining: Improving How Alignment Training Generalizes (NeurIPS 2026 · similarity 0.87)
- Varying Shades of Wrong: Aligning LLMs with Wrong Answers Only (ICLR 2025 · similarity 0.87)
- An Empirical Study on Noisy Data and LLM Pretraining Loss Divergence (NeurIPS 2026 · similarity 0.87)
- Delving into the Reversal Curse: How Far Can Large Language Models Generalize? (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
