AI Paper Atlas

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

Why Self-Distillation Helps and Hurts: Denoising vs. Signal Forgetting

Mingqi Wu, Archer Yi Yang, Qiang Sun

ICML 2026 · Label-efficient and robust learning › Knowledge distillation · approach: Rethinking distillation objectives

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
add more self-distillation iterations on fresh covariates (overparameterized regime)iterative self-distillation, linear regression · Power-law covariance simulation (Sigma_ii=1/i)test MSEMixed
increase initial label noise to sigma = 0.8 versus 0.4self-distillation with ResNet-50 · CIFAR-10test errorImproved
increase training set size nself-distillation with ResNet-50 · CIFAR-10test errorImproved
renew a larger fraction alpha of covariates each iterationself-distillation, linear simulation · Appendix I simulationtest risk dynamicsMixed
filter unlabeled samples aligned with the estimated signal subspace (smaller q)self-distillation, linear simulation · Appendix I simulationminimum test errorImproved
mix a fraction alpha_mix of real noisy labels into each iterationself-distillation, linear simulation · Appendix I simulationlong-term test riskImproved
apply self-distillation when initial teacher has little label noiseself-distillation with ResNet-50 · CIFAR-10 (low noise)test errorHurt

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