Research map / Label-efficient and robust learning / Noisy labels
Intrinsic Gradient Suppression for Label-Noise Prompt Tuning in Vision–Language Models
Jia-yu Li, Jiaxin Qi, Sheng Zhou, Jianqiang Huang, Xian-Sheng Hua
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 cross-entropy with double-softmax cross-entropy | CoOp prompt tuning · Average of 9 datasets, 80% sym noise | accuracy (%): 57.77 -> 74.22 | Improved |
| replace cross-entropy with double-softmax cross-entropy | CoOp prompt tuning · FGVCAircraft, 80% sym noise | accuracy (%): 12.27 -> 29.12 | Improved |
| use double-softmax instead of NLPrompt sample selection plus MAE | prompt tuning with noisy labels · Flowers102, 40% sym noise | accuracy (%): 93.29 -> 91.12 | Hurt |
| use double-softmax instead of NLPrompt | prompt tuning with noisy labels · Caltech101, 90% sym noise | accuracy (%): 87.80 -> 86.68 | Hurt |
| use double-softmax instead of NLPrompt | prompt tuning with noisy labels · DTD, 80% pair noise | accuracy (%): 30.93 -> 33.66 | Improved |
| use double-softmax instead of naive consistent-sample selection (ablation) | prompt tuning with noisy labels · DTD, 80% pair noise | accuracy (%): 5.44 -> 33.66 | Improved |
| use double-softmax instead of LogitNorm (ablation) | prompt tuning with noisy labels · Caltech101, 60% sym noise | accuracy (%): 82.14 -> 95.31 | Improved |
| use double-softmax instead of NCE loss (ablation) | prompt tuning with noisy labels · OxfordPets, 30% pair noise | accuracy (%): 71.91 -> 92.83 | Improved |
| use double-softmax instead of square logit normalization (ablation) | prompt tuning with noisy labels · Caltech101, 30% pair noise | accuracy (%): 81.59 -> 95.85 | Improved |
| use double-softmax instead of LogitClip with a fixed threshold (ablation) | prompt tuning with noisy labels · Caltech101, 60% sym noise | accuracy (%): 49.31 -> 95.31 | Improved |
| add double-softmax to suppress gradients of mislabeled samples | CoOp prompt tuning · Caltech101, 60% sym noise | gradient magnitude on logits | Improved |
| tune LogitClip threshold tau (ablation) | LogitClip · DTD sym noise | accuracy (%): more than 15% drop | Hurt |
Closest papers by meaning
- ZIP: An Efficient Zeroth-order Prompt Tuning for Black-box Vision-Language Models (ICLR 2025 · similarity 0.91)
- Black-Box Test-Time Prompt Tuning for Vision-Language Models (AAAI 2025 · similarity 0.90)
- Understanding and Mitigating Miscalibration in Prompt Tuning for Vision-Language Models (ICML 2025 · similarity 0.89)
- SharpZO: Hybrid Sharpness-Aware Vision Language Model Prompt Tuning via Forward-Only Passes (NeurIPS 2025 · similarity 0.89)
- Adaptive Parameter Selection for Tuning Vision-Language Models (CVPR 2025 · similarity 0.89)
- Nemesis: Normalizing the Soft-prompt Vectors of Vision-Language Models (ICLR 2024 · similarity 0.89)
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.
