AI Paper Atlas

Research map / Label-efficient and robust learning / Recognition and long-tail

GUIDE: Gated Uncertainty-Informed Disentangled Experts for Long-tailed Recognition

Yuan Dong, Zhe Zhao, Liheng Yu, Di Wu, Pengkun Wang

ICLR 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Decoupled and calibrated long-tail training

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 Level 1 competitive specialization (feature decorrelation + JSD) (ablation)entangled multi-expert baseline · CIFAR-100-LT IR=100 (standard schedule)overall top-1 accuracy (%): 45.8 -> 50.4Improved
add Level 2 uncertainty-guided expert refinement (ablation)entangled multi-expert baseline · CIFAR-100-LT IR=100 (standard schedule)overall top-1 accuracy (%): 45.8 -> 51.3Improved
add Level 3 two-timescale updates (ablation)entangled multi-expert baseline · CIFAR-100-LT IR=100 (standard schedule)overall top-1 accuracy (%): 45.8 -> 49.9Improved
add Level 2 on top of Level 1 (ablation)GUIDE Level 1 · CIFAR-100-LT IR=100 (standard schedule)overall top-1 accuracy (%): 50.4 -> 52.8Improved
add Level 3 on top of Levels 1 and 2 (ablation)GUIDE Level 1+2 · CIFAR-100-LT IR=100 (standard schedule)overall top-1 accuracy (%): 52.8 -> 56.4Improved
add only the feature decorrelation loss (ablation)entangled baseline · CIFAR-100-LT IR=100 (standard schedule)overall top-1 accuracy (%): 45.8 -> 47.7Improved
add only the JSD prediction divergence loss (ablation)entangled baseline · CIFAR-100-LT IR=100 (standard schedule)overall top-1 accuracy (%): 45.8 -> 47.5Improved
combine both diversity losses (ablation)entangled baseline · CIFAR-100-LT IR=100 (standard schedule)overall top-1 accuracy (%): 45.8 -> 50.4Improved
use epistemic/aleatoric decomposed gating instead of total uncertainty (ablation)GUIDE DERM gating · CIFAR-100-LT IR=100 (standard schedule)overall top-1 accuracy (%): 54.9 -> 56.4Improved
use decomposed uncertainty gating instead of inverse-frequency static gate (ablation)GUIDE DERM gating · CIFAR-100-LT IR=100 (standard schedule)overall top-1 accuracy (%): 53.6 -> 56.4Improved
use adaptive gating instead of fixed gate 0.5 (ablation)GUIDE DERM gating · CIFAR-100-LT IR=100 (standard schedule)overall top-1 accuracy (%): 52.1 -> 56.4Improved
increase experts from one to three (ablation)GUIDE · CIFAR-100-LT IR=100 (standard schedule)few-shot accuracy: 10.9% absolute gainImproved

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