Research map / Large language models / Language, safety and interpretability
Do Not Merge My Model! Safeguarding Open-Source LLMs Against Unauthorized Model Merging
Qinfeng Li, Miao Pan, Jintao Chen, Fu Teng, Zhiqiang Shen, Ge Su, Hao Peng, Xuhong Zhang
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 MergeBarrier protection to the expert before merging | expert model (Task Arithmetic merge) · GSM8K (LM & Math merge) | merged-model accuracy: 62.17 -> 0.00 | Improved |
| add MergeBarrier protection to the expert before merging | expert model (Task Arithmetic merge) · MBPP (LM & Code merge) | merged-model accuracy: 35.60 -> 0.00 | Improved |
| add MergeBarrier protection to the expert before merging | expert model (TIES merge) · AlpacaEval (LM & Code merge) | merged-model accuracy: 44.62 -> 5.04 | Improved |
| use PaRaMS instead of no protection | expert model · GSM8K (Task Arithmetic, LM & Math merge) | merged-model accuracy: 62.17 -> 59.60 | No clear effect |
| use TaylorMLP instead of no protection | expert model · MBPP (TIES, Three Models merge) | merged-model accuracy: 30.80 -> 24.80 | No clear effect |
| apply MergeBarrier to protected model | protected expert used alone · GSM8K | accuracy: 58.76 -> 57.88 | No clear effect |
| apply MergeBarrier to protected model | protected expert used alone · MBPP | pass@1: 27.80 -> 27.80 | No clear effect |
| apply MergeBarrier to protected model | protected expert used alone · GSM8K | sharpness: 0.57 -> 0.58 | No clear effect |
| add FFN reparameterization to attention-only protection (ablation) | MergeBarrier · average over merge settings | relative mean merged accuracy: 3.79x -> 1.00x | Improved |
| add attention projection to FFN-only protection (ablation) | MergeBarrier · average over merge settings | relative mean merged accuracy: 7.19x -> 1.00x | Improved |
| FFN-only protection versus no protection (ablation) | MergeBarrier (FFN only) · average over merge settings | relative mean merged accuracy: 8.27x -> 7.19x | Mixed |
| add MergeBarrier protection against attacker with fine-tuning | expert model · GSM8K/MBPP with fine-tuning | merged-model performance | Improved |
Closest papers by meaning
- CoreGuard: Safeguarding Foundational Capabilities of LLMs Against Model Stealing in Edge Deployment (NeurIPS 2025 · similarity 0.88)
- Tamper-Resistant Safeguards for Open-Weight LLMs (ICLR 2025 · similarity 0.85)
- MergeBench: A Benchmark for Merging Domain-Specialized LLMs (NeurIPS 2025 · similarity 0.85)
- Single Character Perturbations Break LLM Alignment (AAAI 2025 · similarity 0.84)
- Position: Editing Large Language Models Poses Serious Safety Risks (ICML 2025 · similarity 0.84)
- On Evaluating the Durability of Safeguards for Open-Weight LLMs (ICLR 2025 · similarity 0.84)
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.
