pith. sign in

arxiv: 2605.23254 · v1 · pith:IMU3AOGMnew · submitted 2026-05-22 · 💻 cs.CV

CARE: Class-Adaptive Expert Consensus for Reliable Learning with Long-Tailed Noisy Labels

Mengke Li , Haiquan Ling , Lihao Chen , Yang Lu , Yiqun Zhang , Hui Huang This is my paper

Pith reviewed 2026-05-25 04:41 UTC · model grok-4.3

classification 💻 cs.CV
keywords long-tailed learningnoisy labelsvision-language modelsclass-adaptivelabel rectificationexpert consensus
0
0 comments X

The pith

Class-adaptive consensus across noisy labels, text embeddings and visual features rectifies long-tailed noisy labels more reliably.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces CARE to tackle the combined problems of long-tailed distributions and noisy labels in image classification. It does this by combining three sources of supervision and applying different agreement standards depending on how common each class is. A sympathetic reader would care because standard approaches tend to fail on rare classes while over-correcting common ones, leading to biased models. If the approach works, it provides a way to get better performance without needing more data or complex models.

Core claim

CARE uses a class-adaptive expert consensus mechanism that requires higher agreement among the three sources for tail classes and allows more flexibility for head classes, then aggregates high-confidence predictions to filter noise and recalibrate class distributions.

What carries the argument

The class-adaptive expert consensus mechanism that sets agreement strictness based on class frequency to combine predictions from observed noisy labels, VLM text embeddings, and visual features.

If this is right

  • Improved correction for tail classes without over-regularizing head classes.
  • Consistent outperformance on synthetic and real-world long-tailed noisy benchmarks.
  • Up to 3.0% performance gains over prior methods.
  • Parameter-efficient framework suitable for practical use.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the three sources prove complementary in more settings, the approach could extend to other noisy label scenarios beyond vision.
  • Further gains might come from incorporating additional expert sources like audio or text-only data.
  • Testing on extremely imbalanced distributions could reveal limits of the frequency-based adaptation.

Load-bearing premise

The three supervision sources provide complementary information that can be weighted adaptively by class frequency without introducing additional errors or biases.

What would settle it

Experiments on a dataset where VLM embeddings are corrupted in a way that aligns with the label noise, showing whether the consensus still improves or degrades performance.

Figures

Figures reproduced from arXiv: 2605.23254 by Haiquan Ling, Hui Huang, Lihao Chen, Mengke Li, Yang Lu, Yiqun Zhang.

Figure 2
Figure 2. Figure 2: Overview of CARE. Different colors in the label denote classes, and intensity reflects confidence. ① illustrates expert consensus (TE and BE agree on Top-1 of TE). The Top-1 confidence of TE that is refined by removing unlikely classes is assigned to the consensus class (blue). Top-2 and Top-3 confidences, lacking consensus, are individually accumulated. ② depicts a no-consensus case, where each expert con… view at source ↗
Figure 3
Figure 3. Figure 3: Noise rate dynamics during training of class splits [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Noise Rate Evolution across CARE Components [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗
read the original abstract

Learning from real-world data is frequently hindered by the compound challenge of long-tailed class distributions and noisy annotations. Existing methods partially address these issues but typically ignore the non-uniform impact of label noise across classes, resulting in ineffective correction for tail classes and over-regularization for head classes. To address this issue, we propose Class-Adaptive Rectification with Experts (CARE), a parameter-efficient framework that leverages three complementary supervision sources from vision-language models (VLM): observed noisy labels, VLM text embeddings, and visual features. CARE introduces a class-adaptive expert consensus mechanism that enforces stricter agreement for tail classes and more permissive agreement for head classes based on class frequency. By aggregating high-confidence predictions across these sources, CARE filters unreliable signals and recalibrates class distributions, yielding more reliable rectification under long-tailed distributions. Extensive experiments on both synthetic and real-world benchmarks demonstrate that CARE consistently outperforms state-of-the-art methods, achieving up to 3.0\% performance gains. The source code is available at https://github.com/qwq123-study/CARE.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The manuscript presents CARE, a framework for reliable learning with long-tailed noisy labels. It aggregates high-confidence predictions from observed noisy labels, VLM text embeddings, and visual features using a class-adaptive expert consensus that enforces stricter agreement for tail classes based on class frequency. The method is claimed to filter unreliable signals and recalibrate class distributions, with experiments showing up to 3.0% gains over state-of-the-art on synthetic and real-world benchmarks.

Significance. Should the results prove robust, this work could contribute to handling compound challenges in real-world data by adaptively using VLM sources. The parameter-efficient nature and code release are strengths. However, the small reported gains require strong evidence to establish significance.

major comments (2)
  1. [Abstract] The central performance claim of up to 3.0% gains lacks any details on baselines, statistical tests, error bars, or data splits. This is load-bearing for assessing the effectiveness of the proposed class-adaptive mechanism.
  2. [Abstract] The assumption that VLM sources provide complementary reliable supervision for tail classes without introducing biases is not supported by any mentioned per-class validation or ablation on the frequency-based thresholds versus uniform ones. This directly impacts the validity of the recalibration for long-tailed distributions.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and outline the revisions we will incorporate.

read point-by-point responses

  1. Referee: [Abstract] The central performance claim of up to 3.0% gains lacks any details on baselines, statistical tests, error bars, or data splits. This is load-bearing for assessing the effectiveness of the proposed class-adaptive mechanism.

    Authors: We agree the abstract would benefit from added context. The full manuscript reports results against multiple state-of-the-art baselines across synthetic and real-world long-tailed noisy datasets, with error bars from repeated runs, statistical significance tests, and explicit data-split details in Sections 4 and 5. We will revise the abstract to name the primary baselines and note that these supporting statistics appear in the experiments. revision: yes

  2. Referee: [Abstract] The assumption that VLM sources provide complementary reliable supervision for tail classes without introducing biases is not supported by any mentioned per-class validation or ablation on the frequency-based thresholds versus uniform ones. This directly impacts the validity of the recalibration for long-tailed distributions.

    Authors: The manuscript contains per-class performance breakdowns and an ablation comparing class-frequency-based thresholds against uniform thresholds, showing gains concentrated on tail classes with no evidence of introduced bias. These appear in the experimental analysis. We will revise the abstract to reference these validations and the frequency-based ablation to better substantiate the assumption. revision: yes

Circularity Check

0 steps flagged

No derivation chain; empirical method evaluated on external benchmarks

full rationale

The paper describes an empirical framework (CARE) that aggregates signals from observed labels, VLM text embeddings, and visual features via a class-adaptive consensus rule modulated by class frequency. No equations, fitted parameters renamed as predictions, self-definitional loops, or load-bearing self-citations appear in the provided text. The method is validated through experiments on synthetic and real-world benchmarks, making the central claims falsifiable outside any internal construction. This matches the default expectation of no significant circularity.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The framework rests on the domain assumption that VLMs supply reliable complementary signals and introduces class-specific agreement thresholds as tunable elements.

free parameters (1)
  • class-adaptive agreement thresholds
    Stricter thresholds for tail classes and permissive ones for head classes are chosen based on class frequency and likely require tuning.
axioms (1)
  • domain assumption VLM text embeddings and visual features supply reliable complementary supervision for label rectification
    Invoked as the basis for the three-source aggregation in the framework.

pith-pipeline@v0.9.0 · 5727 in / 1117 out tokens · 32005 ms · 2026-05-25T04:41:53.767401+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

139 extracted references · 139 canonical work pages · 1 internal anchor

  1. [1]

    2025 , pages=

    Li, Yuhang and Li, Zhuying and Jia, Yuheng , booktitle=ICCV, title=. 2025 , pages=

    work page 2025

  2. [2]

    Zhe Zhao and Haibin Wen and Pengkun Wang and Shuang Wang and Zhenkun Wang and Qingfu Zhang and Yang Wang , title =

    work page

  3. [3]

    Zhe Zhao and Pengkun Wang and Haibin Wen and Wei Xu and Song Lai and Qingfu Zhang and Yang Wang , title =

    work page

  4. [4]

    Pengkun Wang and Zhe Zhao and Haibin Wen and Fanfu Wang and Binwu Wang and Qingfu Zhang and Yang Wang , title =

    work page

  5. [5]

    Binwu Wang and Pengkun Wang and Wei Xu and Xu Wang and Yudong Zhang and Kun Wang and Yang Wang , title =

    work page

  6. [6]

    Junnan Li and Dongxu Li and Caiming Xiong and Steven C. H. Hoi , title =

    work page

  7. [7]

    Ilya O. Tolstikhin and Neil Houlsby and Alexander Kolesnikov and Lucas Beyer and Xiaohua Zhai and Thomas Unterthiner and Jessica Yung and Andreas Steiner and Daniel Keysers and Jakob Uszkoreit and Mario Lucic and Alexey Dosovitskiy , title =

    work page

  8. [8]

    Manning , title =

    Jeffrey Pennington and Richard Socher and Christopher D. Manning , title =

    work page

  9. [9]

    Robust training with ensemble consensus , booktitle = ICLR, year =

    Jisoo Lee and Sae. Robust training with ensemble consensus , booktitle = ICLR, year =

    work page

  10. [10]

    Lukas Hoyer and Dengxin Dai and Luc Van Gool , title =

    work page

  11. [11]

    2022 , author =

    Class imbalance in out-of-distribution datasets: Improving the robustness of the TextCNN for the classification of rare cancer types , journal = JBI, volume =. 2022 , author =

    work page 2022

  12. [12]

    O'Connor and Kevin McGuinness , title =

    Paul Albert and Eric Arazo and Tarun Krishna and Noel E. O'Connor and Kevin McGuinness , title =

    work page

  13. [13]

    Lei Feng and Bo An , title =

    work page

  14. [14]

    Adaptive Graph Guided Disambiguation for Partial Label Learning , journal = TPAMI, volume =

    Deng. Adaptive Graph Guided Disambiguation for Partial Label Learning , journal = TPAMI, volume =

    work page

  15. [15]

    Xinyuan Ji and Zhaowei Zhu and Wei Xi and Olga Gadyatskaya and Zilong Song and Yong Cai and Yang Liu , title =

    work page

  16. [16]

    2024 , pages =

    Yexiong Lin and Yu Yao and Tongliang Liu , title =. 2024 , pages =

    work page 2024

  17. [17]

    Learning from Long-Tailed Noisy Data with Sample Selection and Balanced Loss , booktitle = IJCAI, pages =

    Lefan Zhang and Zhang. Learning from Long-Tailed Noisy Data with Sample Selection and Balanced Loss , booktitle = IJCAI, pages =

    work page

  18. [18]

    Kankanhalli , title =

    Junnan Li and Yongkang Wong and Qi Zhao and Mohan S. Kankanhalli , title =

    work page

  19. [19]

    Yisen Wang and Xingjun Ma and Zaiyi Chen and Yuan Luo and Jinfeng Yi and James Bailey , title =

    work page

  20. [20]

    2020 , booktitle = ICML, volume =

    Zheng, Songzhu and Wu, Pengxiang and Goswami, Aman and Goswami, Mayank and Metaxas, Dimitris and Chen, Chao , title =. 2020 , booktitle = ICML, volume =

    work page 2020

  21. [21]

    2019 , pages=

    Probabilistic End-To-End Noise Correction for Learning With Noisy Labels , author=. 2019 , pages=

    work page 2019

  22. [22]

    , title =

    Zhang, Zhilu and Sabuncu, Mert R. , title =. 2018 , booktitle = NIPS, pages =

    work page 2018

  23. [23]

    2018 , pages=

    Tanaka, Daiki and Ikami, Daiki and Yamasaki, Toshihiko and Aizawa, Kiyoharu , booktitle=CVPR, title=. 2018 , pages=

    work page 2018

  24. [24]

    2018 , volume =

    Dimensionality-Driven Learning with Noisy Labels , author =. 2018 , volume =

    work page 2018

  25. [25]

    CleanNet: Transfer Learning for Scalable Image Classifier Training With Label Noise , booktitle = CVPR, pages =

    Kuang. CleanNet: Transfer Learning for Scalable Image Classifier Training With Label Noise , booktitle = CVPR, pages =

    work page

  26. [26]

    Mingcai Chen and Yuntao Du and Wenyu Jiang and Baoming Zhang and Shuai Feng and Yi Xin and Chongjun Wang , title =

    work page

  27. [27]

    Mengmeng Sheng and Zeren Sun and Zhenhuang Cai and Tao Chen and Yichao Zhou and Yazhou Yao , title =

    work page

  28. [28]

    DINOv2: Learning Robust Visual Features without Supervision , journal = TMLR, volume =

    Maxime Oquab and Timoth. DINOv2: Learning Robust Visual Features without Supervision , journal = TMLR, volume =

    work page

  29. [29]

    Zhi-Fan Wu and Tong Wei and Jianwen Jiang and Chaojie Mao and Mingqian Tang and Yufeng Li , title =

    work page

  30. [30]

    Junnan Li and Caiming Xiong and Steven C. H. Hoi , title =

    work page

  31. [31]

    , booktitle=ICCV, title=

    Li, Junnan and Xiong, Caiming and Hoi, Steven C.H. , booktitle=ICCV, title=. 2021 , pages=

    work page 2021

  32. [32]

    Yikai Zhang and Songzhu Zheng and Pengxiang Wu and Mayank Goswami and Chao Chen , title =

    work page

  33. [33]

    Early-Learning Regularization Prevents Memorization of Noisy Labels , volume =

    Liu, Sheng and Niles-Weed, Jonathan and Razavian, Narges and Fernandez-Granda, Carlos , booktitle = NIPS, pages =. Early-Learning Regularization Prevents Memorization of Noisy Labels , volume =

    work page

  34. [34]

    2019 , pages=

    Understanding and Utilizing Deep Neural Networks Trained with Noisy Labels , author=. 2019 , pages=

    work page 2019

  35. [35]

    Lu Jiang and Di Huang and Mason Liu and Weilong Yang , title =

    work page

  36. [36]

    Influence-balanced loss for imbalanced visual classification , author=

    work page

  37. [37]

    Selective-supervised contrastive learning with noisy labels , author=

    work page

  38. [38]

    Robust early-learning: Hindering the memorization of noisy labels , author=

    work page

  39. [39]

    Improving the Accuracy of Learning Example Weights for Imbalance Classification , author=

    work page

  40. [40]

    2022 , volume=

    Adaptformer: Adapting vision transformers for scalable visual recognition , author=. 2022 , volume=

    work page 2022

  41. [41]

    Stochastic Feature Averaging for Learning with Long-Tailed Noisy Labels , author =

    work page

  42. [42]

    2022 , volume =

    Prototype-Anchored Learning for Learning with Imperfect Annotations , author =. 2022 , volume =

    work page 2022

  43. [43]

    Delving into Sample Loss Curve to Embrace Noisy and Imbalanced Data , volume =

    Jiang, Shenwang and Li, Jianan and Wang, Ying and Huang, Bo and Zhang, Zhang and Xu, Tingfa , year =. Delving into Sample Loss Curve to Embrace Noisy and Imbalanced Data , volume =

    work page

  44. [44]

    Prototypical Classifier for Robust Class-Imbalanced Learning

    Wei, Tong and Shi, Jiang-Xin and Li, Yu-Feng and Zhang, Min-Ling. Prototypical Classifier for Robust Class-Imbalanced Learning. 2022

    work page 2022

  45. [45]

    2017 , publisher=

    Faster R-CNN: Towards real-time object detection with region proposal networks , author=. 2017 , publisher=

    work page 2017

  46. [46]

    Learning from massive noisy labeled data for image classification , author=

    work page

  47. [47]

    2018 , pages=

    Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels , author=. 2018 , pages=

    work page 2018

  48. [48]

    Co-teaching: Robust training of deep neural networks with extremely noisy labels , author=

    work page

  49. [49]

    Dividemix: Learning with noisy labels as semi-supervised learning , author=

    work page

  50. [50]

    Identifying mislabeled data using the area under the margin ranking , author =

    work page

  51. [51]

    Unicon: Combating label noise through uniform selection and contrastive learning , author=

    work page

  52. [52]

    Yao, Yazhou and Sun, Zeren and Zhang, Chuanyi and Shen, Fumin and Wu, Qi and Zhang, Jian and Tang, Zhenmin , booktitle=CVPR, pages=. Jo-

    work page

  53. [53]

    Making deep neural networks robust to label noise: A loss correction approach , author=

    work page

  54. [54]

    Using trusted data to train deep networks on labels corrupted by severe noise , author=

    work page

  55. [55]

    Meta-weight-net: Learning an explicit mapping for sample weighting , author=

    work page

  56. [56]

    Dual T: reducing estimation error for transition matrix in label-noise learning , author=

    work page

  57. [57]

    Part-dependent Label Noise: Towards Instance-dependent Label Noise , volume =

    Xia, Xiaobo and Liu, Tongliang and Han, Bo and Wang, Nannan and Gong, Mingming and Liu, Haifeng and Niu, Gang and Tao, Dacheng and Sugiyama, Masashi , booktitle = NIPS, pages =. Part-dependent Label Noise: Towards Instance-dependent Label Noise , volume =

    work page

  58. [58]

    Class-dependent label-noise learning with cycle-consistency regularization , author=

    work page

  59. [59]

    2015 , volume=

    Tongliang Liu and Dacheng Tao , journal=TPAMI, title=. 2015 , volume=

    work page 2015

  60. [60]

    Sample selection with uncertainty of losses for learning with noisy labels , author=

    work page

  61. [61]

    Tong Wei and Jiang-Xin Shi and Wei-Wei Tu and Yu-Feng Li , title =. Front. Comput. Sci. , volume =

    work page

  62. [62]

    2024 , organization=

    Foster adaptivity and balance in learning with noisy labels , author=. 2024 , organization=

    work page 2024

  63. [63]

    Heteroskedastic and imbalanced deep learning with adaptive regularization , author =

    work page

  64. [64]

    Uncertainty-aware learning against label noise on imbalanced datasets , author =

    work page

  65. [65]

    When noisy labels meet long tail dilemmas: A representation calibration method , author=

    work page

  66. [66]

    Label-noise learning with intrinsically long-tailed data , author =

    work page

  67. [67]

    Extracting Clean and Balanced Subset for Noisy Long-tailed Classification , year =

    Li, Zhuo and Zhao, He and Li, Zhen and Liu, Tongliang and Guo, Dandan and Wan, Xiang , journal =. Extracting Clean and Balanced Subset for Noisy Long-tailed Classification , year =

    work page

  68. [68]

    Long-Tail Learning with Foundation Model: Heavy Fine-Tuning Hurts , author=

    work page

  69. [69]

    Class-balanced loss based on effective number of samples , author=

    work page

  70. [70]

    Khan and Munawar Hayat and Mohammed Bennamoun and Ferdous Ahmed Sohel and Roberto Togneri , title =

    Salman H. Khan and Munawar Hayat and Mohammed Bennamoun and Ferdous Ahmed Sohel and Roberto Togneri , title =. 2018 , doi =

    work page 2018

  71. [71]

    Learning imbalanced datasets with label-distribution-aware margin loss , author=

    work page

  72. [72]

    2021 , pages =

    Wang, Jiaqi and Zhang, Wenwei and Zang, Yuhang and Cao, Yuhang and Pang, Jiangmiao and Gong, Tao and Chen, Kai and Liu, Ziwei and Loy, Chen Change and Lin, Dahua , title =. 2021 , pages =

    work page 2021

  73. [73]

    Cao, Dong and Zhu, Xiangyu and Huang, Xingyu and Guo, Jianzhu and Lei, Zhen , title =

    work page

  74. [74]

    2021 , volume=

    Zhang, Wanping and Chen, Yongru and Yang, Wenming and Wang, Guijin and Xue, Jing-Hao and Liao, Qingmin , journal=TNNLS, title=. 2021 , volume=

    work page 2021

  75. [75]

    Tan, Jingru and Wang, Changbao and Li, Buyu and Li, Quanquan and Ouyang, Wanli and Yin, Changqing and Yan, Junjie , title =

    work page

  76. [76]

    2021 , pages =

    Hong, Youngkyu and Han, Seungju and Choi, Kwanghee and Seo, Seokjun and Kim, Beomsu and Chang, Buru , title =. 2021 , pages =

    work page 2021

  77. [77]

    Ren, Mengye and Zeng, Wenyuan and Yang, Bin and Urtasun, Raquel , title =

    work page

  78. [78]

    2020 , volume=

    Huang, Chen and Li, Yining and Loy, Chen Change and Tang, Xiaoou , journal=PAMI, title=. 2020 , volume=

    work page 2020

  79. [79]

    Focal Loss for Dense Object Detection , journal = TPAMI, volume =

    Tsung. Focal Loss for Dense Object Detection , journal = TPAMI, volume =

    work page

  80. [80]

    and Zoph, Barret and Man

    Cubuk, Ekin D. and Zoph, Barret and Man. AutoAugment: Learning Augmentation Strategies From Data , booktitle = CVPR, pages =

    work page

Showing first 80 references.