AI Paper Atlas

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

ICML 2026 · Architectures and efficiency › Pruning and efficient tuning · approach: Low-rank compression and merging · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
train adapter with Trap2 objective (instead of plain LoRA) before 8-way TA mergingLoRA adapter, CLIP ViT-B/32 · 8-way TA merging (Full)avg accuracy (lower = better protection): 48.273 -> 23.121Improved
train adapter with Trap2 before TA mergingLoRA adapter, CLIP ViT-L/14 · 8-way TA merging (Full)avg accuracy (lower = better protection): 62.914 -> 33.619Improved
train adapter with Trap2 before TA mergingLoRA adapter, CLIP ConvNeXt · 8-way TA merging (Full)avg accuracy (lower = better protection): 49.203 -> 14.719Improved
apply PaRaMS (adapter-space) post-hoc protectionLoRA adapter, CLIP ViT-B/32 · 8-way TA merging (Full)avg accuracy (lower = better protection): 48.273 -> 48.290No clear effect
apply Merge-Lock post-hoc protection to adaptersLoRA adapter, CLIP ViT-B/32 · standalone avg 8 tasksaccuracy: 88.119 -> 6.254Hurt
apply PaRaMS refit variant to adaptersLoRA adapter, CLIP ViT-L/14 · standalone avg 8 tasksaccuracy: 92.403 -> 83.887Hurt
train with Trap2 instead of plain fine-tuningLoRA adapter, CLIP ViT-B/32 · standalone avg 8 tasksaccuracy: 88.119 -> 88.132No clear effect
train with Trap2 under full fine-tuningCLIP ViT-B/32 full FT · standalone avg 8 tasksaccuracy: 87.598 -> 87.372No clear effect
train with Trap2 under full fine-tuning before TA mergingCLIP ViT-B/32 full FT · full-model TA mergingavg accuracy (lower = better protection): 49.963 -> 28.475Improved
train with Trap2 against data-dependent RegMean mergingLoRA adapter, CLIP ViT-B/32 · RegMean mergeavg accuracy (lower = better protection): 49.107 -> 9.480Improved
train with Trap2 against Chain-of-MergesLoRA adapter, CLIP ViT-B/32 · CoM mergeavg accuracy (lower = better protection): 64.776 -> 32.331Improved
train with Trap2 against ProDistill recoveryLoRA adapter, CLIP ViT-B/32 · ProDistill recoveryavg accuracy (lower = better protection): 73.823 -> 49.153Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy