Research map / Label-efficient and robust learning / Recognition and long-tail
CURE: Consistency-under-Unified Semantic Regularization for Generalized Category Discovery
Yuwei Bian, Shidong Wang, Haofeng Zhang
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add FT, RSE and LSKL components to the baseline setup (ablation) | CURE · CUB-200 | ACC (All): 70.4 -> 75.2 | Improved |
| add FT, RSE and LSKL components (ablation) | CURE · CIFAR-100 | ACC (All): 81.2 -> 85.2 | Improved |
| add supervised fine-tuning with prototype alignment in stage I (ablation) | CURE · CIFAR-100 | ACC (New): 72.5 -> 80.0 | Improved |
| use hierarchical clustering instead of K-means for centroids (ablation) | CURE · CUB-200 | ACC (All): 73.4 -> 75.2 | Improved |
| use spectral instead of hierarchical clustering (ablation) | CURE · CUB-200 vs CIFAR-100 | ACC (All): 75.2 -> 76.7 on CUB; 85.7 -> 85.6 on CIFAR-100 | Mixed |
| increase virtual sampling multiplier t to 1.3 (ablation) | CURE · CUB-200 | Old-New gap: 5.6% -> 0.1% | Improved |
| increase virtual sampling multiplier t to 1.2 (ablation) | CURE · CIFAR-100 | Old-New gap: 7.3% -> 1.3% | Improved |
| lower energy threshold epsilon from 0.90 to 0.80 (ablation) | CURE · CIFAR-100 | ACC (All): 85.5 -> 84.1 | Hurt |
| set epsilon 0.95 instead of 1.00 (ablation) | CURE · CUB-200 | ACC (All): 73.6 -> 74.6 | Improved |
| add stage II semantic expansion on top of stage I (ablation) | CURE · ImageNet-100 | ACC (All): 83.2 -> 87.0 | Improved |
| use DINOv2 backbone instead of DINO | CURE · CIFAR-100 | ACC (All): 85.7 -> 89.1 | Improved |
| add the symmetric divergence consistency loss (ablation) | CURE · CUB-200 and CIFAR-100 | ACC | Improved |
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.
- DebGCD — DebGCD: Debiased Learning with Distribution Guidance for Generalized Category Discovery (ICLR 2025) · better on 6 of 6 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
- Consistent Supervised-Unsupervised Alignment for Generalized Category Discovery (NeurIPS 2025 · similarity 0.91)
- SEAL: Semantic-Aware Hierarchical Learning for Generalized Category Discovery (NeurIPS 2025 · similarity 0.89)
- Transfer and Alignment Network for Generalized Category Discovery (AAAI 2024 · similarity 0.88)
- HiLo: A Learning Framework for Generalized Category Discovery Robust to Domain Shifts (ICLR 2025 · similarity 0.88)
- HIDISC: A Hyperbolic Framework for Domain Generalization with Generalized Category Discovery (NeurIPS 2025 · similarity 0.88)
- Towards Understanding Parametric Generalized Category Discovery on Graphs (ICML 2025 · similarity 0.88)
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.
