Research map / Architectures and efficiency / Pruning and efficient tuning
Making Models Unmergeable via Scaling-Sensitive Loss Landscape
Minwoo Jang, Hoyoung Kim, Jabin Koo, Jungseul Ok
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 |
|---|---|---|---|
| train adapter with Trap2 objective (instead of plain LoRA) before 8-way TA merging | LoRA adapter, CLIP ViT-B/32 · 8-way TA merging (Full) | avg accuracy (lower = better protection): 48.273 -> 23.121 | Improved |
| train adapter with Trap2 before TA merging | LoRA adapter, CLIP ViT-L/14 · 8-way TA merging (Full) | avg accuracy (lower = better protection): 62.914 -> 33.619 | Improved |
| train adapter with Trap2 before TA merging | LoRA adapter, CLIP ConvNeXt · 8-way TA merging (Full) | avg accuracy (lower = better protection): 49.203 -> 14.719 | Improved |
| apply PaRaMS (adapter-space) post-hoc protection | LoRA adapter, CLIP ViT-B/32 · 8-way TA merging (Full) | avg accuracy (lower = better protection): 48.273 -> 48.290 | No clear effect |
| apply Merge-Lock post-hoc protection to adapters | LoRA adapter, CLIP ViT-B/32 · standalone avg 8 tasks | accuracy: 88.119 -> 6.254 | Hurt |
| apply PaRaMS refit variant to adapters | LoRA adapter, CLIP ViT-L/14 · standalone avg 8 tasks | accuracy: 92.403 -> 83.887 | Hurt |
| train with Trap2 instead of plain fine-tuning | LoRA adapter, CLIP ViT-B/32 · standalone avg 8 tasks | accuracy: 88.119 -> 88.132 | No clear effect |
| train with Trap2 under full fine-tuning | CLIP ViT-B/32 full FT · standalone avg 8 tasks | accuracy: 87.598 -> 87.372 | No clear effect |
| train with Trap2 under full fine-tuning before TA merging | CLIP ViT-B/32 full FT · full-model TA merging | avg accuracy (lower = better protection): 49.963 -> 28.475 | Improved |
| train with Trap2 against data-dependent RegMean merging | LoRA adapter, CLIP ViT-B/32 · RegMean merge | avg accuracy (lower = better protection): 49.107 -> 9.480 | Improved |
| train with Trap2 against Chain-of-Merges | LoRA adapter, CLIP ViT-B/32 · CoM merge | avg accuracy (lower = better protection): 64.776 -> 32.331 | Improved |
| train with Trap2 against ProDistill recovery | LoRA adapter, CLIP ViT-B/32 · ProDistill recovery | avg accuracy (lower = better protection): 73.823 -> 49.153 | Improved |
Closest papers by meaning
- CoreGuard: Safeguarding Foundational Capabilities of LLMs Against Model Stealing in Edge Deployment (NeurIPS 2025 · similarity 0.84)
- On Evaluating the Durability of Safeguards for Open-Weight LLMs (ICLR 2025 · similarity 0.83)
- Mitigating the Backdoor Effect for Multi-Task Model Merging via Safety-Aware Subspace (ICLR 2025 · similarity 0.83)
- MGit: A Model Versioning and Management System (ICML 2024 · similarity 0.82)
- AnchorRep: Defending LLMs Against Cross-Model Adversarial Transfer via Representation Repulsion (NeurIPS 2026 · similarity 0.82)
- Representation Noising: A Defence Mechanism Against Harmful Finetuning (NeurIPS 2024 · similarity 0.82)
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.
