Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Generalized Logit Adjustment: Calibrating Fine-tuned Models by Removing Label Bias in Foundation Models
Beier Zhu, Kaihua Tang, Qianru Sun, Hanwang 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 |
|---|---|---|---|
| apply GLA on top of E2E/WiSE-FT | CLIP ViT-B/16 E2E · ImageNet | accuracy: 81.3 -> 82.8 | Improved |
| GLA instead of WiSE-FT ensembling | CLIP ViT-B/16 · ImageNet | accuracy: 81.7 -> 82.8 | Improved |
| WiSE-FT ensembling without bias handling | E2E · ImageNet tail classes | accuracy: 77.3 -> 76.4 | Hurt |
| GLA on tail classes | E2E · ImageNet tail classes | accuracy: 77.3 -> 78.8 | Improved |
| remove label bias from zero-shot model (ablation) | CLIP ViT-B/16 zero-shot · ImageNet | accuracy: 68.8 -> 69.3 | Improved |
| transfer pi_p from ViT-B/32 to ViT-L/14 (ablation) | zero-shot CLIP ViT-L/14 · ImageNet | accuracy: 75.6 -> 76.3 | Improved |
| GLA vs prompt tuning baselines under shift | prompt tuning · ImageNet-A | accuracy: 21.3 -> 23.0 | Improved |
| prompt tuning baseline vs zero-shot CLIP under shift | CoOp · ImageNet-Sketch | accuracy: 35.5 -> 32.5 | Hurt |
| debias zero-shot using estimated pi_p (ablation) | CLIP-RN50 zero-shot · 11 datasets average | accuracy: 59.1 -> 60.7 | Improved |
| naive ensembling instead of GLA ensemble (ablation) | fine-tuned models · EuroSAT, SUN, Caltech, UCF, Aircraft | accuracy | Hurt |
| GLA instead of BALLAD | E2E CLIP-RN50 · ImageNet-LT | accuracy: 67.2 -> 68.2 | Improved |
| GLA instead of balanced softmax | prompt tuning · ImageNet-LT | accuracy: 63.5 -> 67.8 | Improved |
Closest papers by meaning
- LEAD: Exploring Logit Space Evolution for Model Selection (CVPR 2024 · similarity 0.86)
- Rethinking the Bias of Foundation Model under Long-tailed Distribution (ICML 2025 · similarity 0.86)
- AMU-Tuning: Effective Logit Bias for CLIP-based Few-shot Learning (CVPR 2024 · similarity 0.86)
- Class and Attribute-Aware Logit Adjustment for Generalized Long-Tail Learning (AAAI 2025 · similarity 0.85)
- Understanding Overadaptation in Supervised Fine-Tuning: The Role of Ensemble Methods (ICML 2025 · similarity 0.85)
- LT-Soups: Bridging Head and Tail Classes via Subsampled Model Soups (NeurIPS 2025 · similarity 0.85)
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.
