Research map / Label-efficient and robust learning / Noisy labels
NoiseGPT: Label Noise Detection and Rectification through Probability Curvature
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes NoiseGPT, which uses an MLLM (MMICL) as a zero-shot expert to detect and rectify label noise via a token-wise mixture-of-feature perturbation and an in-context discrepancy score based on probability curvature.
Problem
Label noise in image datasets degrades deep network training; existing learning-with-noisy-labels methods need memorization-based clean/noisy separation or strong dataset assumptions and do not provide cleaned datasets across scenarios.
Approach
Observes that clean and noisy image-label pairs lie on probability curvatures of different smoothness in an MLLM; perturbs image features by token-wise mixing with exemplar features and scores the In-Context Discrepancy (normalized drop in True-probability) to detect noise, then rectifies labels by picking the CLIP top-C candidate with highest ICD score.
Key results
- NoiseGPT reaches AUROC 0.9253 on ILSVRC12 Sym. 40% and 0.8935 on WebVision Sym. 40%, and 0.7899 on CIFAR-100N Noisy.
- On CIFAR-10 Sym. 80%, NoiseGPT+M-correction reaches 92.6% accuracy versus 87.6% for M-correction alone (about 5 points), and detection F1 of 96.07% vs 93.49% for DivideMix*.
- On CIFAR-10 Sym. 90%, NoiseGPT+M-correction reaches 91.5% versus 68.7% for M-correction, a gain of over 20 points.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| clean the training data with NoiseGPT before M-correction | M-correction · CIFAR-10 Sym. 90% | accuracy: 68.7 -> 91.5 | Helped |
| clean the training data with NoiseGPT before Pro-Mix | Pro-Mix · CIFAR-10 Sym. 90% | accuracy: 75.0 -> 89.6 | Helped |
| clean the training data with NoiseGPT before M-correction | M-correction · CIFAR-100 Sym. 80% | accuracy: 47.3 -> 58.1 | Helped |
| clean the training data with NoiseGPT before Pro-Mix | Pro-Mix · CIFAR-100 Sym. 80% | accuracy: 55.4 -> 63.9 | Helped |
| clean the training data with NoiseGPT before DivideMix | DivideMix · Webvision validation | accuracy: 77.32 -> 78.10 | Helped |
| increase the number of perturbations (ablation) | NoiseGPT · noise rectification experiments | rectification performance | Helped |
| decrease the MoF weight, increasing perturbation strength (ablation) | NoiseGPT · CIFAR-10 Sym. 50%, CIFAR-100 Sym. 50%, Webvision Sym. 40%, ILSVRC Sym. 40% | output Softmax probability | Mixed |
Surprising findings
- Providing a set of class candidates and asking which is ground truth performs poorly for MLLMs, so the method restricts outputs to True/False binary answers.
- Rectification accuracy on WebVision and ILSVRC12 (about 43%) is much lower than detection AUROC, indicating label correction remains hard.
Limitations
- Performance is bounded by the capability of the underlying MLLM and its vision encoder bottleneck.
- ICD distributions depend on the MLLM's instruction-following and answer confidence, and prompt settings matter.
- No statistical significance or error bars reported due to limited compute; detection shows class-wise biases.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 CIFAR-10N CIFAR-100N WebVision ILSVRC2012 mini-WebVision
Benchmarks: CIFAR-10 CIFAR-100 CIFAR-10N CIFAR-100N WebVision ILSVRC2012
Metrics: AUROC label correction accuracy precision recall F1 classification accuracy
Baselines: M-correction Pro-Mix DivideMix Proto-Mix F-correction Co-teaching+ Mixup P-correction
Builds on: MMICL CLIP DetectGPT-style probability curvature M-correction Pro-Mix
Closest papers by meaning
- Handling Label Noise via Instance-Level Difficulty Modeling and Dynamic Optimization (NeurIPS 2025 · similarity 0.89)
- Delving into Noisy Label Detection with Clean Data (ICML 2023 · similarity 0.89)
- Learning with Feature-Dependent Label Noise: A Progressive Approach (ICLR 2021 · similarity 0.89)
- Does label smoothing mitigate label noise? (ICML 2020 · similarity 0.89)
- Mitigating Label Noise through Data Ambiguation (AAAI 2024 · similarity 0.88)
- On the Role of Label Noise in the Feature Learning Process (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.