AI Paper Atlas

Research map / Label-efficient and robust learning / Test-time adaptation

Blocking the Leakage: Manifold-Aware Gradient Projection for Long-Horizon Test-Time Adaptation

Haoyu Xiong, Chengchao Wang, ZhongQiang Wang, He Huang, Qiuxia Yang, Zhengpeng Zhao, Yuanyuan Pu

ICML 2026 · Label-efficient and robust learning › Test-time adaptation · approach: Realistic and stable test-time adaptation

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
project entropy gradients onto the orthogonal complement of a tracked low-rank subspaceTENT-style entropy minimization · ImageNet-C long-horizon revisits (40 rounds)accuracy: TENT 37.51 -> 0.46 vs MGP 41.90 -> 42.96Improved
add MGP to the stream with only misclassified samples (blind-spot)entropy-based TTA · ImageNet-K, ImageNet-R blind-spotaccuracy: IN-K 27.01, IN-R 32.85 vs baselines near 0%Improved
tighten entropy gating threshold to 0.1 on TENT (ablation)TENT · ImageNet-C (ViT-B/16) long-horizonrounds before collapse: collapse delayed to 11 rounds, not preventedMixed
loosen entropy gating threshold to 0.4 or above on TENT (ablation)TENT · ImageNet-C (ViT-B/16) long-horizonrounds before collapse: collapse within 3 roundsHurt
project out the protected subspace from spurious-sample gradients (ablation)entropy minimization · ImageNet-Caccuracy over timeImproved
project reliable-sample gradients onto the residual space (ablation)entropy minimization with reliable samples · ImageNet-CaccuracyNo clear effect
increase refresh interval K from 100 to 1000 (ablation)MGP · ImageNet-C long-horizonfinal accuracy change: -1.89%Hurt
inject strong rank-1 anisotropy into residual updates (kappa=100) (ablation)MGP · ImageNet-C long-horizonfinal accuracy change: -1.80%Hurt
use automatic MP bulk-edge rank selection instead of fixed rank (ablation)MGP · ImageNet-C long-horizonaccuracyImproved
shuffle domain order every round (ablation)MGP · ImageNet-C (ViT-B/16)accuracy: 53.13 -> 55.26No clear effect
abrupt shift from ImageNet-C to imbalanced ImageNet-K (ablation)MGP · two-stage shift (ViT-B/16)accuracy: 36.68 -> 44.36Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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