acceptodds
Under review as a conference paper at ICLR 2027

Retrieval-Calibrated One-Step Categorical Denoising for Learning with Severe Label Noise

Abstract

Learning with severe label noise requires a model to recover useful supervision without prematurely committing to potentially incorrect labels. Recent label-diffusion methods address this problem by progressively denoising corrupted label states, but it remains unclear which part of the diffusion formulation is actually responsible for their effectiveness: iterative reverse generation, or learning from label states with systematically varying reliability. We investigate this question through RC-Cat, a retrieval-calibrated categorical denoising framework that separates corruption-aware training from iterative generative inference. RC-Cat first converts each noisy annotation into a sparse candidate distribution using reliability-calibrated neighborhood evidence, thereby preserving uncertainty over the unknown clean class. It then corrupts sampled categorical label states across different noise levels and trains a shared timestep- conditioned denoiser to reconstruct the candidate distribution. Controlled experiments reveal a clear mechanism. On CIFAR-100 with severe PMD-70 noise, introducing categorical corruption improves accuracy from 60.29% to 64.78%, while exposing the corruption timestep further raises it to 71.29%. In contrast, adding the exact categorical posterior regularizer produces no measurable improvement, and iterative reverse inference does not outperform direct prediction from the terminal corruption level. When the representation and candidate supervision are held fixed, RC-Cat exceeds a parameter-matched soft-target classifier by 5.96 percentage points; the advantage remains 2.53 points when externally pretrained CLIP features are replaced by a target-data-only SimCLR representation. Improvements also extend to composite synthetic noise and CIFAR-100N, although they become smaller on Animal-10N and negligible on Clothing1M. These results suggest that the useful inductive bias of label diffusion can arise primarily from corruption-aware learning: exposing a classifier to label states of varying reliability and explicitly informing it of their corruption level. From this perspective, iterative label generation is not necessary to obtain the observed denoising benefit, enabling a one-step categorical formulation that is substantially smaller and faster than a protocol-matched iterative diffusion baseline.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Diffusion model, Label learning, Label generation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Vs3WOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  1. The paper proposes RC-Cat, a one-step categorical label diffusion framework for noisy-label classification. It avoids expensive iterative reverse-chain sampling, achieving inference throughput comparable to a simple MLP baseline while retaining the benefits of label diffusion training. A key conceptual insight is validated: iterative reverse-chain inference is not required to obtain the denoising benefit. The performance gain mainly originates from training on corrupted label states with timestep conditioning, rather than multi-step sampling at test time.
  2. The paper designs a carefully controlled matched baseline (parameter-matched MLP). It isolates the performance gain brought by the corruption-conditioning training paradigm, removing confounding factors from representation and retrieval.
  3. Extensive comparisons are conducted against both classic noisy-label learning methods (Co-teaching+, GCE, PLC, etc.) and state-of-the-art label diffusion approaches (DLD, LRA). The method achieves strong performance especially under severe synthetic label noise (PMD-70).
Critical strengths
  1. The paper proposes and validates a key insight: denoising gains can be obtained without iterative reverse-chain inference. The performance improvement mainly stems from training on corrupted label states with timestep conditioning, rather than multi-step sampling at test time.
  2. The authors carefully design a controlled baseline (parameter-matched MLP) for comparative experiments.
  3. The mathematical formulation in the method section is clear and easy to follow; Figures 1 and 2 are also intuitive.
  4. The authors have provided partial code and commit to releasing the full code as open source.
Critical weaknesses
  1. The abstract contains multiple intermediate staged accuracy values (60.29%→64.78%→71.29%). While these figures are important for the experimental section, they constitute redundant details in the abstract. They consume the limited word budget of the abstract and distract readers from the high-level scientific contributions.
  2. Nearly half of the current abstract describes experimental observations. For a mechanism-focused paper, the abstract should prioritize highlighting: open research problems remaining in label diffusion; the high-level design of RC-Cat; the core insight that iterative reverse inference is not mandatory; the overall advantages and applicability boundaries of the one-step scheme. Specific numbers from stepwise ablation studies are recommended to be moved to the experimental section in the main text.
  3. Table 2 only enumerates published values from existing literature rather than results reproduced under a unified experimental setup. Traditional label-noise-robust methods such as Co-Teaching, GCE and PLC directly adopt results reported in their original papers. Differences in pre-trained features, data augmentation, training epochs and optimizers mean these values can only serve as references and cannot be used for rigorous comparisons.
  4. Ablation on inference hyper-parameters: full sweep over K (the number of Monte Carlo samples). Currently only K=4 is validated to be effective, without sensitivity analysis for hyper-parameter K.
  • examine how accuracy saturates with increasing K;
  • observe the magnitude of variance reduction;
  • analyze scenarios where K=1 is sufficient and scenarios requiring a large K.
  1. The main text includes the passage: On Clothing1M, the nearly identical performance of RC-Cat and the matched classifier may additionally reflect the different large-scale regime: with roughly one million training examples and only 14 classes, strong frozen representations and candidate supervision may already make the classification problem sufficiently constrained that corruption-aware training provides little additional benefit. We therefore view the benefit of RC-Cat as noise- and data-regime dependent rather than uniform across noisy-label settings. Although Appendix A2 contains several numerical results, the relationships among these data remain unclear. please specify the exact performance metrics. Explain why no performance gain is observed and verify this conjecture. Further validate the performance of the proposed method on large-scale real-world noisy-label datasets.

Totally, experiments are enough for supporting claims in the paper

Questions for authors

to see above weaknesses.

Reviewer_ZNJKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes RC-Cat, a categorical denoising framework for learning with severe label noise. It combines retrieval-based soft label construction with timestep-conditioned categorical corruption and one-step inference, avoiding iterative reverse denoising. Experiments on synthetic and real-world noisy-label datasets show improvements over matched soft-target baselines, while ablations suggest that corruption-aware training, rather than iterative inference, is the main source of performance gains.

Critical strengths
  1. The paper is well motivated and easy to follow. The proposed method is closely connected to the research question, and the empirical findings provide reasonable support for the design choices.

  2. The experiments include controlled comparisons and ablation studies to analyze the effectiveness of different components. The method also offers a reasonable accuracy-efficiency trade-off compared with iterative diffusion baselines.

Critical weaknesses
  1. Limited methodological novelty. The proposed method mainly combines existing techniques, including retrieval-based label correction, categorical corruption, and timestep conditioning. While the empirical findings are useful, the methodological contribution appears incremental. It is also unclear what unique advantages the diffusion formulation offers over simpler label augmentation or conditional classification methods.

  2. Questionable necessity of the diffusion framework. Since the task is essentially classification with fixed retrieval-based soft labels, it is unclear why a diffusion framework is needed in the first place. The authors emphasize one-step inference as a key contribution, but single-step prediction is already standard in classification. This raises questions about whether the proposed framework introduces unnecessary complexity rather than offering a meaningful methodological improvement.

  3. Insufficient evaluation on real-world label noise. The main improvements are observed under synthetic noise, while the gains on real-world noisy-label datasets are relatively small. Moreover, widely used benchmarks such as WebVision are missing, despite being evaluated in closely related prior work. This raises concerns about the effectiveness and generalizability of the proposed method in more challenging real-world settings.

Questions for authors
  1. The improvements on real-world noisy-label datasets are relatively small. Could the authors provide additional results on more challenging benchmarks, such as Mini-WebVision, to better demonstrate the effectiveness of RC-Cat under real-world label noise?

  2. The ablation studies show substantial gains from timestep conditioning. Have the authors considered using random or shuffled timesteps as a control to verify that the improvements come from accurate corruption-level information rather than additional conditioning?

Reviewer_JZfFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a label-noise learning method, RC-Cat, that simplifies and reduces the computational requirements of diffusion-based label-noise approaches. The method shows that the iterative reverse process used by label-diffusion methods does not bring a significant accuracy improvement and can be removed for improved efficiency. The authors also propose a more interpretable alternative for the label-state corruption mechanism and a Monte Carlo-based inference setup to average predictions over several terminal steps. The experimental section provides basic baselines and ablation studies to show the contribution of different elements of RC-Cat.

Critical strengths
  • The authors identify an interesting problem: the computational complexity of diffusion-based label-noise learning setups. The proposed approach reduces computational costs and is grounded in weaknesses observed in existing approaches.
  • The paper provides valuable insights into the mechanisms behind the approach and shows that, for RC-Cat, reverse-chain inference brings little improvement while adding computational cost.
  • The results show a large reduction in checkpoint size and a reasonable increase in throughput. This trade-off comes with a small decrease in accuracy.
Critical weaknesses
  • Based on Table 3, the results are good in terms of throughput and checkpoint storage size. However, these improvements come at the cost of accuracy. If the contribution is mainly on the efficiency side, the text should be more oriented towards this and the introduction should organize the contributions accordingly. In that case, efficiency should be measured in terms of computational cost and throughput. It would be valuable to report end-to-end FLOPs/image and batch-one latency alongside GPU memory usage during inference.
  • Along the same lines, the authors should clarify whether the reported performance/efficiency improvements account for all elements of the approach, including the image encoder and denoiser. The additional computation introduced by the KK predictions at inference time should also be measured and accounted for in the benchmarking.
  • The interaction between the model and the label-noise distribution is not addressed in depth. It seems that the model performs well under different types of noise, but the paper should include a discussion of the effect that different noise distributions have on the method. For instance, Table 1 seems to suggest that the approach handles asymmetric noise more easily than symmetric noise. This should be clarified, and further observations and discussion around this behaviour would be valuable.
  • The quality of qq should be further analyzed and discussed. This seems to be a key element that may vary depending on the image encoder, dataset, and type of noise present in the original dataset. A sensitivity analysis varying candidate-target quality or retrieval settings would strengthen this point. Please consider further discussing the dependence on qq quality and how errors in qq affect RC-Cat.
  • The experimental section is limited in terms of literature context. Table 2 reports results that provide an approximate indication of where RC-Cat stands, but they do not allow for a fair performance comparison. While the authors already label this as contextual, the paper would benefit from a matched comparison or from additional discussion/guidance so that these results are more interpretable.
  • The generalization of the approach to more complex datasets and noise distributions should be further assessed. The initial results on real-world noise are promising, but the efficiency-accuracy trade-off is not detailed for this type of noise. Reporting accuracy, throughput, and inference cost alongside each other for Animal-10N and Clothing-1M would clarify when RC-Cat offers a useful trade-off.
Questions for authors

See weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.