Research map / Large language models / Instruction and fine-tuning
In-Training Defenses Against Emergent Misalignment in Language Models
David Kaczér, Magnus Jørgenvåg, Clemens Vetter, Esha Afzal, Robin Haselhorst, Lucie Flek, Florian Mai
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 |
|---|---|---|---|
| add KL-divergence regularization toward aligned model | SFT on EM datasets, Qwen2.5-7B · EM Code (general) | misaligned rate: 4.01 -> 0.38 | Improved |
| add LDIFS feature-space l2 regularization | SFT on EM datasets · EM datasets (general) | misaligned rate: 25.29 -> 26.75 (Legal) | No clear effect |
| add preventive persona vector steering | SFT on EM datasets, Qwen2.5-7B · EM Legal (general) | misaligned rate: 25.29 -> 1.00 | Improved |
| add inoculation prompting | SFT on EM datasets, Qwen2.5-7B · EM Legal (general) | misaligned rate: 25.29 -> 16.46 | Mixed |
| interleave 5% automatically selected safety data (Interleaving++) | SFT on EM datasets, Qwen2.5-7B · EM Legal (general) | misaligned rate: 25.29 -> 0.79 | Improved |
| select interleaving data by perplexity gap and filter refusals (vs random) (ablation) | Interleaving · EM general domain | incoherence vs dataset size | Improved |
| add KL-divergence regularization | SFT on OpSwap · OpSwap Tier 1 | exact match: 30.00 -> 0.00 | Hurt |
| add KL-divergence regularization | SFT on FoQA · FoQA | exact match: 41.60 -> 44.55 | Improved |
| add persona vector steering | GRPO on GSM8K, Qwen2.5-3B · GSM8K | accuracy: 90.66 -> 38.33 | Hurt |
| add 20% interleaved SFT data | GRPO on GSM8K, Qwen2.5-3B · GSM8K | accuracy: 90.66 -> 93.66 | Improved |
| add Persona Vectors or KL to reduce general misalignment | EM fine-tuning · EM Medical in-domain | in-domain misalignment: 51.73 -> 29.67 (Persona Vectors) | Hurt |
| add Interleaving++ to preserve narrow misalignment | EM fine-tuning · EM in-domain | in-domain misalignment: 51.60 -> 52.80 (Code) | No clear effect |
Closest papers by meaning
- Antidote: Post-fine-tuning Safety Alignment for Large Language Models against Harmful Fine-tuning Attack (ICML 2025 · similarity 0.91)
- Panacea: Mitigating Harmful Fine-tuning for Large Language Models via Post-fine-tuning Perturbation (NeurIPS 2025 · similarity 0.90)
- Alignment Dynamics in LLM Fine-Tuning (NeurIPS 2026 · similarity 0.90)
- Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (ICLR 2024 · similarity 0.89)
- Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models (NeurIPS 2024 · similarity 0.89)
- Booster: Tackling Harmful Fine-tuning for Large Language Models via Attenuating Harmful Perturbation (ICLR 2025 · similarity 0.89)
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.
