AI Paper Atlas

Research map / Label-efficient and robust learning / Knowledge distillation

DisTaC: Conditioning Task Vectors via Distillation for Robust Model Merging

Kotaro Yoshida, Yuji Naraki, Takafumi Horie, Ryotaro Shimizu, Ioannis Mitliagkas, Hiroki Naganuma

ICLR 2026 · Label-efficient and robust learning › Knowledge distillation · approach: Few-shot and metric distillation

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
apply DisTaC (norm and confidence conditioning via distillation) before mergingEMR-Merging · 8 vision tasks, Low Confidenceabsolute accuracy: 27.4 -> 92.3 (ViT-L-14)Improved
apply DisTaC before mergingTSVM · 8 vision tasks, Low Confidenceabsolute accuracy: 60.7 -> 81.5 (ViT-B-32)Improved
apply DisTaC before mergingWUDI-Merging · 8 vision tasks, Norm Mismatchabsolute accuracy: 49.2 -> 84.4 (ViT-B-32)Improved
apply DisTaC before mergingTask arithmetic · 8 vision tasks, Norm Mismatchabsolute accuracy: 63.6 -> 70.0 (ViT-B-32)Improved
apply DisTaC before merging under low confidenceIso-CTS · 8 vision tasks, Low Confidenceabsolute accuracy: 72.5 -> 69.0 (ViT-B-32); 80.8 -> 86.1 (ViT-L-14)Mixed
apply DisTaC under norm mismatch for Iso-CTS ViT-L-14Iso-CTS · 8 vision tasks, Norm Mismatchabsolute accuracy: 90.8 -> 90.1No clear effect
train one task with learning rate 10^-4 instead of 10^-5 (norm disparity)model merging methods · 8 vision tasks, CLIP ViT-B-32normalized accuracy: up to 14% dropHurt
fine-tune source models with label smoothing 0.1model merging methods · 8 vision tasks, CLIP ViT-B-32normalized accuracy: up to 24% dropHurt
increase label smoothing strength during source fine-tuning (ablation)model merging · 8 vision tasks, CLIP ViT-B-32normalized accuracyHurt
scale task vector down (kappa < 1) (ablation)task vector · 8 vision tasks, CLIP ViT-B-32test accuracyNo clear effect
stretch task vector up to kappa=3.0 (ablation)task vector · 8 vision tasks, CLIP ViT-B-32test accuracyHurt
distill with student temperature 10 and teacher temperature 1 to raise confidence (ablation)source models · 8 vision tasks, CLIP ViT-B-32prediction entropyImproved

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