Research map / Vision-language and multimodal / CLIP and zero-shot
SoC: Semantic Orthogonal Calibration for Test-Time Prompt Tuning
Leo Fillioux, Omprakash Chakraborty, Ismail Ben Ayed, Paul-Henry Cournède, Stergios Christodoulidis, Maria Vakalopoulou, Jose Dolz
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 |
|---|---|---|---|
| replace full orthogonality penalty (O-TPT) with Huber regularizer (SoC) | TPT · Average of 11 datasets (ViT-L/14) | ECE: 7.7 -> 5.4 | Improved |
| replace full orthogonality penalty with Huber regularizer | TPT · Average of 11 datasets (ViT-L/14) | accuracy: 71.4 -> 72.3 | Improved |
| replace full orthogonality penalty with Huber regularizer | TPT · EuroSAT | ECE: 17.7 -> 3.2 | Improved |
| replace full orthogonality penalty with Huber regularizer | TPT · ImageNet | ECE: 5.5 -> 7.2 | Hurt |
| replace full orthogonality penalty with Huber regularizer | TPT · Average of 11 datasets (ViT-B/16) | ECE: 4.8 -> 4.3 | Improved |
| replace full orthogonality penalty with Huber regularizer | TPT · ImageNet variants average | ECE: 9.9 -> 8.4 | Improved |
| replace full orthogonality penalty with Huber regularizer | TPT · ImageNet variants average | accuracy: 71.3 -> 71.3 | No clear effect |
| add a second gradient update on text prompts (ablation) | SoC vs O-TPT · Average of 11 datasets | ECE: SoC degraded nearly 23%; O-TPT 39% | Hurt |
| replace full orthogonality penalty with Huber regularizer under two gradient updates | TPT · Average of 11 datasets | ECE: 10.8 -> 6.6 | Improved |
| replace full orthogonality penalty with Huber regularizer with CoOp 4-shot prompts | TPT+CoOp · Average of 11 datasets | accuracy: 76.6 -> 78.0 | Improved |
| normalize similarities with min-max (S1) instead of dividing by max (S2) (ablation) | SoC · Average of 8 datasets | accuracy and ECE: accuracy 70.7 -> 70.3; ECE 7.3 -> 5.3 | Mixed |
| normalize similarities with min-max (S1) instead of subtracting min (S3) (ablation) | SoC · Average of 8 datasets | accuracy: 69.2 -> 70.3 | 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.
- O-TPT — O-TPT: Orthogonality Constraints for Calibrating Test-time Prompt Tuning in Vision-Language Models (CVPR 2025) · better on 8 of 10 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
- O-TPT: Orthogonality Constraints for Calibrating Test-time Prompt Tuning in Vision-Language Models (CVPR 2025 · similarity 0.93)
- C-TPT: Calibrated Test-Time Prompt Tuning for Vision-Language Models via Text Feature Dispersion (ICLR 2024 · similarity 0.88)
- Open-Vocabulary Calibration for Fine-tuned CLIP (ICML 2024 · similarity 0.87)
- PARC: A Quantitative Framework Uncovering the Symmetries within Vision Language Models (CVPR 2025 · similarity 0.87)
- Nemesis: Normalizing the Soft-prompt Vectors of Vision-Language Models (ICLR 2024 · similarity 0.87)
- Amend to Alignment: Decoupled Prompt Tuning for Mitigating Spurious Correlation in Vision-Language Models (ICML 2024 · similarity 0.86)
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.
