acceptodds
Under review as a conference paper at ICLR 2027

SmoothOperator: Enhancing Representations for Fine-grained Open-set Recognition via Modulated Label Smoothing

Abstract

Open Set Recognition (OSR) aims to enable models to accurately classify known classes while rejecting samples from unseen classes. A key challenge in OSR lies in the inability to model the unbounded distribution of unknown classes during training, often leading to the misclassification of samples from these classes. Rather than modeling unknowns, recent work shapes the feature space so that known classes are compact and well separated, and spherical representation learning methods have achieved strong results this way. Label smoothing has been identified as one of the key drivers of this success, yet it applies the same coefficient to every training sample, regardless of how well each sample is already embedded. We show that the spherical representation learning objectives used in OSR share a single alignment–uniformity structure in which labels enter only through the alignment term. Label smoothing therefore acts as an alignment dial, and a fixed coefficient sets this dial to the same value for every sample. We propose a plug-in, SmoothOperator (SmoothOP), which sets the smoothing coefficient of each sample from its prominence, an embedding-space signal measuring how clearly the sample's own class stands out against its strongest competing class. Our method integrates into four existing spherical representation learning methods at minimal training overhead. SmoothOP assigns strong smoothing to samples with high prominence, which reduces their alignment and relaxes their pull. Conversely, samples with low prominence receive reduced smoothing, which translates to increased alignment, pulling them to the class center. Empirically, SmoothOP yields more compact class clusters and, despite modifying only the alignment term, better dispersed ones, which together improve OSR performance. On the Semantic Shift Benchmark, SmoothOP-augmented variants generally outperform their base objectives across datasets, degrees of semantic shift, and OSR post-processors, with gains of up to 4.7% in AUROC, OSCR, and closed-set accuracy.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Representation Learning, Open-set Recognition, Label Smoothing

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_4ohGOpenReview ↗
Rating
2
Confidence
4(+1)
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on fine-grained open-set recognition from the perspective of label smoothing in spherical representation learning. By analyzing four existing representation learning objectives, the authors identify a shared alignment–uniformity structure and propose SmoothOP, which adaptively adjusts the smoothing coefficient for each sample based on its prominence, estimated from classifier confidence margins. Experiments on the SSB datasets demonstrate improvements in fine-grained OSR performance.

Critical strengths
  • The paper is well structured and generally clearly written, with a logical presentation of the motivation and methodology. Especially the progression from the alignment uniformity analysis to the proposed adaptive smoothing mechanism.

  • The proposed method is intuitively motivated and can be easily incorporated into existing representation learning frameworks without substantial architectural modifications.

  • The experimental evaluation is relatively comprehensive, covering multiple datasets, and contrastive learning objects. Additional ablation studies and feature geometry analyses provide further insights into the proposed approach.

Critical weaknesses
  • As a standalone submission, a) although this paper is generally well written, too many methodological details and design choices rely heavily on the SpHOR[1], requiring readers to frequently refer to that work to fully understand the proposed framework, both in the main text and the Appendix. b) In my view, the paper is closer to an incremental investigation of the label-smoothing strategy adopted in existing representation learning methods, particularly SpHOR, building on the observations of [2]. The main methodological contribution is to replace a fixed smoothing coefficient with a sample dependent one.

  • The alignment–uniformity decomposition has already been extensively studied in prior literature. In particular, in OSR field [3] provides a closely related analysis in their subsection “Rethinking the Properties of Contrastive Learning,” where they explicitly show that label-dependent similarity weights affect only the alignment term, while the uniformity term remains independent of them. This observation substantially overlaps with the theoretical motivation of the present paper, raising concerns about the novelty of its theoretical contribution.

  • While the motivation is reasonable, its conceptual novelty appears limited, especially considering existing research on instance-dependent adaptive label smoothing. In particular, [4] already adapt the smoothing coefficient per sample according to the model’s predictive uncertainty. The paper does not sufficiently establish what constitutes a substantial methodological advance beyond these prior works.

  • The empirical improvement is not sufficiently consistent to support some of the paper's stronger claims.

[1] SpHOR: A representation learning perspective on open-set recognition for identifying unknown classes in deep neural networks. CVPR 2026.

[2] Open-set recognition: A good closed-set classifier is all you need. ICLR 2022.

[3] Contrastive open set recognition. AAAI 2023

[4] Adaptive label smoothing with self-knowledge in natural language generation. EMNLP 2022.

Questions for authors

In this section, my main concerns are related to several details of the experimental evaluation.

  • The experiments cover only three fine-grained datasets from the Semantic Shift Benchmark [2], while the ImageNet benchmark included in the original SSB is omitted. Could the authors explain this choice? I would also like to see evaluations on conventional OSR benchmarks, such as MNIST, SVHN, CIFAR, and TinyImageNet, to better assess the generalizability of the proposed method beyond fine-grained recognition.

  • In Table 3, applying SmoothOP to ConOSR on CUB reduces closed-set accuracy from 75.89% to 75.42%. Similarly, Table 4 shows a slight deterioration in intra-class compactness under the same setting. Could the authors provide further analysis of this degradation?

  • I would like the authors to discuss OpenAUC[1]. Furthermore, I am not fully convinced by the justification for excluding OOSA in Appendix A. In particular, the authors argue against using surrogate unknown validation data for threshold selection, yet such data are used to tune the proposed method's hyperparameter α\alpha. Could the authors further justify their position and clarify this distinction?

[5] Openauc: Towards auc-oriented open-set recognition. NeruIPS 2022.

8 comments
AuthorsOpenReview ↗

Official Response to Reviewer 4ohG (Part 1/5)

Overall Response: We thank the reviewer for their comments and questions. We grouped the critical weaknesses and questions by topic and answered each group below. We have also revised the paper. All edits are in dark blue, and we refer to them by line number.


CW1(a): Paper is not self-contained.

We agree with this comment. We have revised the paper and made it more self-contained. We clarified that the paper uses the Supervised Contrastive Learning framework (with other methods such as SpHOR/ConOSR) (Lines 183-189) and provided a generalized equation for it. (“A common template”, Lines 244-264).


CW1(b): Incremental investigation and lacks methodological contribution.

We respectfully disagree that our work is an incremental investigation of the label smoothing used in SpHOR. ConOSR/SpHOR use label smoothing as a fixed training choice; they do not ask how it acts on a representation objective, or how it should vary across samples. Our work answers both questions, and the answers are not specific to SpHOR. We clarify this below, together with the differences from [4].

(1) Not specific to SpHOR. Our proposed method integrates into four objectives, SupCON, ConOSR, SpHOR, and SpHOR-Lite (SpHOR without Mixup), with no change to the architecture or the objective and at negligible cost (Table 7), and it improves their OSR performance in most settings (Table 2).

(2) A design principle. An adaptive smoothing method is a control mechanism: it uses an input signal to vary the smoothing coefficient per sample, and the coefficient in turn controls a property of the learned model. To design it, one must know what the coefficient controls, and which input should set it. Our proposed method contributes along both axes. Most adaptive label smoothing methods are designed for classifiers [4, 9, 10]. In a classifier, the label is the target distribution that the predicted class probabilities must match, and smoothing acts as entropy regularization of the output, which mitigates overconfidence [7]. In a representation objective, the label plays a different role: it decides which samples are positives and which are negatives for each anchor (Eq. 4). Building on the alignment–uniformity decomposition, which we unify across these objectives (Prop. 1), we show that label smoothing acts as an alignment dial (Prop. 2; details in our response to CW 2). We therefore propose this alignment view of label smoothing as a design principle for adaptive label smoothing in representation learning.

(3) A signal designed for OSR. Once we know that smoothing controls alignment, the input must answer: which samples need more alignment? Adaptive smoothing methods answer such questions with a task-specific signal that encodes the inductive bias of their task (See CW 3 answer). The inductive bias of OSR is geometric: known classes must form compact, well-separated clusters, so that the remaining space is reserved for unknowns [11]. We therefore propose prominence, a signal that adjusts the alignment of each sample to produce this geometry. To our knowledge, it is the first adaptive label smoothing method designed for OSR. We explain the design and its differences from [4] in our response to CW 3.

(4) Mixup. A Mixup sample has two classes, so a per-sample coefficient is not defined for it. To our knowledge, no prior adaptive label smoothing method defines one; ours is the first. We compute the prominence of the mixed sample as the Mixup-weighted combination of the prominence of its two constituent samples, and apply the resulting single coefficient once to the mixed target (Eqs. 10–11). In our ablation, we show this form of blending outperforms blending the two smoothing coefficients directly in 11 of 12 settings (App. D.3, Table 12).

[4] Lee, D., et al, Adaptive label smoothing with self-knowledge in natural language generation. EMNLP'22.

[7] Pereyra, G., et al, Regularizing neural networks by penalizing confident output distributions. ICLRW'17.

[9] Bae, J., et al, CaliMatch: Adaptive calibration for improving safe semi-supervised learning. ICCV'25.

[10] Bahavan, T. T. N., et al, FedLAS: Feature-modulated bidirectional label smoothing for neural network calibration. ECCV'26.

[11] Scheirer, W. J., et al, Toward open set recognition. TPAMI'13.

(continued in Part 2/5)

AuthorsOpenReview ↗

Official Response to Reviewer 4ohG (Part 2/5)

CW2: Theoretical Contribution.

We agree that the decomposition itself is not new, and we thank the reviewer for this pointer. We revised the paper to make this clear. The decomposition builds on Wang & Isola [5], and ConOSR [3] and SpHOR each show it for their own objective. We extend and show that it holds across four OSR objectives in one template (Prop. 1, Table 1), and we rewrote contribution statement 1 (lines 94–100) to say this. Our theoretical contribution is what follows from the decomposition, and it motivates our proposed method, SmoothOP.

(1) The extension to label smoothing. ConOSR [3] shows that label-dependent similarity weights enter only the alignment term. However, it does not link this result to label smoothing: ConOSR and SpHOR both use a fixed smoothing coefficient, so the question of how the coefficient changes the objective does not arise. In Prop. 2 (with Props. 4–5 in App. E), we show that increasing the smoothing coefficient moves target mass from positives to negatives monotonically and by a quantified amount, while the uniformity term stays unchanged. This is what makes the per-sample coefficient of our proposed method interpretable: each value of εi\varepsilon_i has a known, one-directional effect on that sample’s alignment. This is the basis of the design principle in point 2.

(2) A theoretical lens for the role of label smoothing, which gives a design principle. *Labels play a different role in classifiers and in representation learning objectives, so label smoothing must also be interpreted differently (see also our response to CW 1(b) and 3). In a classifier, the label is the target of the predicted output, and label smoothing [6] is commonly interpreted as entropy regularization of that output [7], which reduces overconfidence and improves calibration [8]. With CC classes, label smoothing replaces the one-hot label yy with y~=(1−ε)y+εu\tilde{y} = (1-\varepsilon) y + \varepsilon u, where u=1C1u = \frac{1}{C}\mathbf{1} is the uniform distribution and ε\varepsilon is the smoothing coefficient. For a classifier with predicted class probabilities pp, the loss splits into two terms:

CE(y~,p)=(1−ε)CE(y,p)+εKL(u∥p)+εlog⁡C\mathrm{CE}(\tilde{y}, p) = (1-\varepsilon) \mathrm{CE}(y, p) + \varepsilon \mathrm{KL}(u \parallel p) + \varepsilon \log C

The first term fits the label. The second term penalizes confidence: a larger ε\varepsilon pushes pp toward higher entropy, i.e., less overconfidence. Existing adaptive methods follow this output-confidence view [4]. In a representation learning objective, the label instead sets the relation between each anchor and its positives and negatives (Eq. 4), and the goal is the geometry of the representation space, not a calibrated output. Props. 1–2 show that, in this setting, increasing smoothing reduces alignment, rather than regularizing output entropy. We therefore propose this alignment view of label smoothing as a design principle for adaptive smoothing in representation learning objectives. It turns the design into a concrete question: which samples need a stronger pull toward their class? Our proposed method answers it with prominence: samples with high prominence receive more smoothing, and samples with low prominence receive less smoothing and a stronger pull toward their class. (revised in lines 80–86) This is our theoretical contribution, and we now state it explicitly in the paper (lines 74-79). We also added a Related work section on classifier-based adaptive label smoothing methods. (Lines 134-139)*

[3] Xu, B., et al, Contrastive open set recognition. AAAI'23.

[4] Lee, D., et al, Adaptive label smoothing with self-knowledge in natural language generation. EMNLP'22.

[5] Wang, T., et al, Understanding contrastive representation learning through alignment and uniformity on the hypersphere. ICML'20.

[6] Szegedy, C., et al, Rethinking the Inception architecture for computer vision. CVPR'16.

[7] Pereyra, G., et al, Regularizing neural networks by penalizing confident output distributions. ICLRW'17.

[8] Müller, R., et al, When does label smoothing help? NeurIPS'19.

(continued in Part 3/5)

AuthorsOpenReview ↗

Official Response to Reviewer 4ohG (Part 3/5)

CW3: Conceptual Novelty of Instance dependent label smoothing.

We agree that instance-adaptive label smoothing exists, and we compare with [4] directly. At a high level, most adaptive methods, including [4] and ours, can be described as varying the smoothing coefficient per sample. The substance lies in the design decisions:

(1) A different setting and framework. [4] adapts the target of a classifier and follows the output-confidence view: it treats low output entropy as overconfidence. Our proposed method adapts the label in a representation objective, where smoothing controls alignment (see our responses to CW 1(b) and 2).

(2) Adaptive smoothing is task-specific, so rules do not transfer between tasks. Each adaptive label smoothing method rests on a design principle built for its own task, and its signal encodes the inductive bias of that task. CaliMatch uses validation accuracy per confidence level and is designed for semi-supervised learning, where unlabeled data are available; FedLAS uses the feature norm to calibrate output probabilities, not to reject unknowns; and [4] uses prediction entropy to counter overconfidence in text generation. A signal that is meaningful under one framework is not necessarily meaningful under another, so these rules cannot be assumed to transfer to OSR. The inductive bias of OSR is geometric: known classes must form compact, well-separated clusters, so that the remaining space is reserved for unknowns [11]. We therefore propose prominence, a signal that sets the alignment of each sample to produce this geometry (Sec. 4). In fine-grained OSR, unknown classes are visually close to known ones, so the regions between neighbouring known classes are the most at risk. Prominence, the margin between a sample’s own class and its strongest competing class, acts exactly on these regions: it gives samples near a class boundary less smoothing and a stronger pull.

Why prominence, and not entropy. Entropy measures how spread the output is, and it does not use the ground-truth label. It therefore cannot tell whether a confident prediction is correct: a confidently correct sample and a confidently misclassified sample both have low entropy, and both receive near-maximal smoothing under ALS [4]. Prominence uses the ground-truth label. The confidently misclassified sample has low prominence, so it receives weak smoothing and a stronger pull toward its own class, while the confidently correct sample receives strong smoothing.

Evidence. App. C.4 (Table 9) replaces only the rule that sets the coefficient with that of [4] or FedLAS, and keeps everything else identical. Our proposed method gives the best average OSCR on both splits for both objectives. We use the non-mixup variants because ALS/FedLAS doesn’t define a Mixup rule.

Average OSCR (%) over CUB, Aircraft and Cars and four post-hoc scores; mean over 3 seeds. Best in bold, second best underlined.

ObjectiveSplitLSALS [4]FedLAS [10]Ours
SupConEasy71.0471.97‾\underline{71.97}70.7572.03
SupConHard65.2765.96‾\underline{65.96}65.7866.32
SpHOR-LiteEasy72.45‾\underline{72.45}71.1971.7472.91
SpHOR-LiteHard66.89‾\underline{66.89}65.7065.4067.18

(3) Extending it to Mixup. (Refer CW 1(b))

[4] Lee, D., et al, Adaptive label smoothing with self-knowledge in natural language generation. EMNLP'22.

[9] Bae, J., et al, CaliMatch: Adaptive Calibration for Improving Safe Semi-supervised Learning. ICCV'25.

[10] Bahavan, T. T. N., et al, FedLAS: Feature-Modulated Bidirectional Label Smoothing for Neural Network Calibration. ECCV'26.

[11] Scheirer, W. J., et al, Toward open set recognition. TPAMI'13.


CW4: Empirical results.

We revised the wording of these claims so that each one precisely reflects the empirical results. The changes are in dark blue in the revised manuscript (e.g. Table 2 Caption, Lines 500-502).

(continued in Part 4/5)

Reviewer_BfBNOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The author proposed a label smoothing method for each sample individually for fine-grained OSR. The evaluations demonstrate its effectiveness compared to some OSR methods.

Critical strengths
  1. The paper is, in general, complete in structure and writing.

  2. The evaluations are sound and can support the advantages of the proposed method.

Critical weaknesses
  1. Minor Comment on Paper Structure: The abstract is unusually long and detailed, whereas the conclusion is very brief (only a few sentences). A more standard academic presentation would benefit from a concise abstract and a more thorough conclusion summarizing key takeaways and potential limitations.

  2. I cannot agree with some claims in the paper:

    a. (line 050-059) "earlier OSR literature has followed two main approaches...", I think one category of methods is missing, which is the discriminative models like MSP.

    b. (line 166-169) "Concretely, OSR methods based on representation learning follow a three-stage pipeline. The first two stages are inherited from, and theoretically grounded in, self-supervised learning(Grill et al., 2020; Chen et al., 2020b)." I think this claim is not complete. Although self-supervised learning is often applied to OSR, it is hard to say that OSR methods based on representation learning are inherited from and grounded in SSL.

  3. It is not stated in the manuscript why this SmoothOperator is useful specifically for fine-grained OSR.

  4. The term "representation learning" is probably too broad to be used in this work. The method is, however, built on top of supervised contrastive learning and its variants. But still, the authors have not explained why the proposed method is important for fine-grained OSR.

  5. The reasons for how the prominence is designed are unclear.

  6. Comparisons with other famous and StoA OSR methods are missing in Table 2.

Questions for authors
  1. Following 4 in the Weaknesses: I cannot see where the representation learning framework in Section 4.1 comes from. I think there is no "standard" or "conventional" framework for representation learning. Maybe supervised contrastive learning for OSR is a better term than representation learning here.

  2. Following 6 in the Weaknesses: Can the method work for normal OSR and other applications?

  3. Following 3 in the Weaknesses: Please clarify the connection between SmoothOperator and fine-grained OSR.

2 comments
AuthorsOpenReview ↗

Official Response to Reviewer BfBN (Part 1/2)

Overall Response: We thank the reviewer for their comments and questions. We grouped the critical weaknesses and questions by topic and answered each group below. We have also revised the paper. All edits are in dark blue, and we refer to them by line number.


CW1: Abstract, Conclusion Structure.

We agree. We expanded the conclusion to summarize the key takeaways and state the limitations (Lines 525-539), and shortened the abstract.


CW2(a): Missing methods

We agree. We added discriminative methods as a third category in the introduction (Lines 50–52)

CW2(b): Incomplete claim related to SSL

We agree. We made this claim more precise in the methodology section (Lines 184–196). The first two stages of these OSR methods follow the supervised contrastive learning framework, which extends self-supervised contrastive learning to labeled data.


CW3 + Q3: Why Fine-grained?

As suggested by the reviewer, we explicitly state this connection in the Conclusion (Lines 529-539), and investigate it in Appendix C.5.

SmoothOP derives each sample’s degree of smoothing from the sample’s prominence, which we defined in the paper. However, if the prominences of all samples converge towards the same value, every sample receives the same smoothing coefficient and SmoothOP degenerates to constant label smoothing. Thus, it is effective only while the prominences take different values, because only then can it rank samples by ambiguity and give ambiguous samples a stronger pull toward their positives.

By nature, the uiu_i of all samples move toward −1-1 (near-full confidence) as the model becomes well trained (Fig. 8). What differs between datasets is how fast this happens. On coarse-grained datasets, most training samples saturate early, while on fine-grained datasets, they approach −1-1 slowly and most samples remain unsaturated for much of training (Appendix C.5, Fig. 6). By mid-training, 46–57% of CIFAR-10 and Tiny ImageNet samples are converged to −1-1, against 0–3% for CUB, 15–27% for Stanford Cars and 28–29% for FGVC-Aircraft (Appendix C.5, Fig. 6). Training with constant smoothing gives nearly identical curves, so this behavior comes from the data and the model, not from SmoothOP.

In the feature space, the coarse-grained datasets form tight, well-separated clusters of saturated samples, while CUB stays spread over the unsaturated range (Fig. 7). This is consistent with the structural properties of fine-grained datasets, such as high inter-class similarity and few images per class. As a result, SmoothOP gives no consistent gain on coarse-grained datasets (Table 10), which have lower inter-class similarity and more images per class.


CW4 + Q1: Clarifying Representation Learning

We agree that “representation learning” is too broad. We now use “supervised contrastive learning” in methodology (184-196) and no longer call the framework “standard” or “conventional”. Section 4 now states that the framework is that of Khosla et al. (2020) (Lines 183–189), and we give a generalized equation for it (Lines 244–257). We explain the connection to fine-grained OSR in our answer to Critical Weakness 3.

(continued in Part 2/2)

AuthorsOpenReview ↗

Official Response to Reviewer BfBN (Part 2/2)

CW5: Design of prominence.

We revised the Introduction (lines 80–86, 68-73) so that each design choice follows from one chain of reasoning:

(1) End state. We first identify the end state (goal) of the algorithm. Following the open-space risk framework of Scheirer et al. (2013), known classes must form compact, well-separated clusters, so that the remaining space stays free for unknowns.

(2) Design principle. By Props. 1–2, label smoothing acts only on alignment: the smoothing coefficient of a sample sets how strongly that sample is pulled toward its own class. Adaptive smoothing therefore needs an input that decides which samples need a stronger pull.

(3) Where the geometry breaks. We then identify how the design principle can achieve the end-state geometry, by examining where this geometry breaks. It breaks at samples that lie close to their strongest competing class, which is the other known class whose cluster is nearest to them. For example, in FGVC-Aircraft, a Boeing 737-800 image lies near the 737-700 cluster, because the two variants differ mainly in fuselage length. Such samples sit at the edge of their own cluster. They make the cluster less compact and narrow the gap between the two clusters. We included this analogy in Lines 80–86.

(4) What the signal must measure. From points 1–3, the signal must measure how close each sample is to its strongest competing class. Samples that are close need a stronger pull (less smoothing). Samples that are already clearly separated need less pull (more smoothing).

(5) Prominence. We therefore define prominence as the margin between the predicted probability of the sample’s own class and that of its strongest competing class (Def. 3). Each part of this definition follows from the points above:

(i) The strongest competitor, not all classes: the geometry breaks against the nearest cluster, not against the average of all classes.

(ii) The ground-truth label: prominence uses the label of the sample, so a misclassified sample has low prominence and gets a stronger pull.

(iii) A gradient-detached observer: the signal reads the embedding but does not change it, so it does not interfere with the objective.

(6) The mapping. The sigmoid mapping (Eq. 9) is bounded, so every smoothed target stays a valid distribution. It is monotone, so more prominence always gives more smoothing. At α=0\alpha = 0, it gives the base method back (Prop. 3).

(1) Scheirer, W. J., et al, Toward open set recognition. TPAMI'13.


CW6 + Q2: Comparisons and other applications.

We added recent state-of-the-art training based OSR methods to Table 2 as representative direct comparisons: BackMix (TPAMI 2025), DEF (AAAI 2025), MEDAF (AAAI 2024) and NegSched/PosSched (ICLR 2026). All use the same architecture, training and splits as ours. On OSCR, our primary metric, the best result is a SmoothOP-augmented method in 5 of the 6 settings.

When applied to normal OSR benchmarks, it does not consistently improve or degrade the performance. We evaluated SmoothOP on the normal OSR benchmarks MNIST, SVHN, CIFAR-10 and Tiny ImageNet with four objectives (Table 10). This follows from the property explained in our answer to Critical Weakness 3: on coarse-grained data, the prominences of most samples converge early, and SmoothOP degenerates to constant label smoothing (Appendix C.5). We state this limitation in Sec. 7. Refer to our answer in CW3+Q3 for details.

Other applications. SmoothOP also improves closed-set classification in most fine-grained settings (Table 3), so its effect is not limited to fine-grained open-set recognition. We have not tested other tasks, but we expect it to help in near-OOD detection and generalized category discovery, which use similar objectives on similar fine-grained data.

(1) Wang, Y., et al, BackMix: Regularizing Open Set Recognition by Removing Underlying Fore-Background Priors. TPAMI'25.

(2) Wang, Y., et al, Exploring diverse representations for open set recognition. AAAI'24.

(3) Kundu, A. K., et al, Boosting open set recognition performance through modulated representation learning. ICLR'26.

(4) Li, C., et al, Unlocking better closed-set alignment based on neural collapse for open-set recognition. AAAI'25.

Reviewer_KLobOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SmoothOP for fine-grained open-set recognition: an auxiliary classifier computes the predicted-probability margin between a sample’s true class and its strongest competing class; a Sigmoid function dynamically adjusts each sample’s label-smoothing strength, and the method is integrated into four spherical representation-learning objectives. The authors report closed-set classification and open-set recognition results on CUB, FGVC-Aircraft, and Stanford Cars, with multiple scoring methods and ablation studies.

Critical strengths
  1. The problem has practical value. The method is simple, can be readily integrated into multiple existing representation-learning objectives, and adds little computational overhead.
  2. Experiments cover three fine-grained datasets, Easy/Hard unknown classes, and four scoring functions; the appendix also examines batch-mean smoothing and reversed assignment, a better design than comparing only with fixed smoothing.
Critical weaknesses
  1. Novelty is insufficiently established. This paper mainly adjusts the label smoothing strength dynamically based on sample predictions, but does not sufficiently clarify its unique novelty. Similar ideas have been explored in prior work. For example, CaliMatch (ICCV 2025) also adopts confidence-based adaptive label smoothing. However, the paper lacks sufficient discussion of and comparison with relevant prior work and recent representative methods. In addition, FedLAS (ECCV 2026) has proposed a plug-in mechanism that modulates label-smoothing coefficients using per-sample signals, and its formulation is similar to that of this paper. Since FedLAS was published recently, a direct comparison is not required in principle; however, I still recommend reading and referring to it to better articulate the novelty.
  2. The method’s effectiveness is insufficiently validated. The paper uses predictions from an auxiliary classifier to assign different label-smoothing strengths to samples of different difficulty, but it has not sufficiently shown that this adjustment produces the claimed improvements in feature space. The intra-class compactness metric in the main text may be affected by changes in feature-vector magnitude, so it cannot directly support the proposed mechanism. More importantly, Appendix Table 8 shows that, compared with using only the mean smoothing coefficient of each batch, the full method improves mean OSCR by only 0.18 and 0.05 percentage points on Easy and Hard, respectively, with no improvement in some settings. Therefore, the existing experiments do not sufficiently establish the independent contribution or necessity of per-sample adjustment.
  3. The paper contains multiple inconsistencies in notation, wording, and descriptions of experimental settings. The authors should standardize the notation and descriptions of experimental settings and thoroughly check the paper’s presentation.

[1] Bae et al. CaliMatch: Adaptive Calibration for Improving Safe Semi-supervised Learning. ICCV 2025.

[2] FedLAS: Feature-Modulated Bidirectional Label Smoothing for Neural Network Calibration. ECCV 2026; arXiv:2606.28654.

Questions for authors
  1. Compared with existing adaptive smoothing work such as CaliMatch (ICCV 2025), what is new about the proposed mechanism? Although the paper includes SpHOR (2026) as a baseline, its main experiments are still limited to adding the module to a few existing representation-learning objectives and lack direct comparisons with recent related methods. Please add representative direct comparison experiments to strengthen the evidence for the method’s novelty.
  2. Do SmoothOP and the baselines use exactly the same Observer/stage-two classification-head initialization, auxiliary supervision, and model-selection budget?
  3. Equations (7)–(8) only guarantee 0<εᵢ<2ε. When the base ε>0.5, εᵢ may exceed 1. How can the smoothed labels always remain valid probability distributions? Should ε≤0.5 be explicitly required?
  4. The paper also contains some basic presentation issues. For example, Eq. (3) on page 4 uses λ for the Mixup weight, while line 368 on page 7 also uses λ for the base smoothing coefficient; should these be distinguished? Stage 2 on page 5 calls the d-dimensional encoded feature f, although f was previously defined as the normalized projected feature; there are also typos such as “from from” (page 2, line 89) and “a a random permutation” (page 4, line 208);
3 comments
AuthorsOpenReview ↗

**Official Response to Reviewer KLob (Part 1/3)**

Overall Response: We thank the reviewer for their comments and questions. We grouped the critical weaknesses and questions by topic and answered each group below. We have also revised the paper. All edits are in dark blue, and we refer to them by line number.


CW1 + Q1(A) - Difference from CaliMatch.

We highlighted the unique novelty of our method over recently published methods such as CaliMatch/FedLAS more explicitly via some additions in the abstract, contributions (93-101), related work (Lines 134-139), methodology (Lines 259 - 264) in the following manner.

(1) Adaptive label smoothing for representation learning, not for a classifier (Lines 259 - 264). Methods such as CaliMatch/FedLAS are designed for classifiers. In a classifier, the label identifies the class and smoothing softens the target labels that the output must predict. In representation learning methods such as supervised contrastive learning, no output predicts the label, and the label instead sets the relations between samples by deciding which elements of the contrast set are positives and which are negatives for each anchor. Adaptively smoothing the label of an anchor therefore softens these relations, so that the attraction to its positives weakens and a small attraction to its negatives appears, while the uniformity between samples does not change (Props. 1–2).

(2) A different design framework (Lines 94-99). Following from the previous point, classifier-based adaptive smoothing is often designed from the view of label smoothing as an entropy regularizer on the output (e.g FedLAS). In our paper, we proposed the alignment–uniformity view of representation learning as the design principle for adaptive smoothing. In this view, the per-sample coefficient becomes a per-sample control of alignment, a property which classifiers don't have. To our knowledge, our proposed SmoothOP is the first method that adapts the label-smoothing coefficient per sample as a plug-in for existing supervised contrastive objectives.

(3) A signal based on the inductive bias of OSR. (Lines 134-139), Each adaptive method uses a signal that suits its own task, and this signal is often not available or not relevant in another task. CaliMatch uses the validation accuracy per confidence level and is designed for semi-supervised learning with unlabeled data, which OSR does not have. FedLAS uses the feature norm to calibrate output probabilities, not to reject unknowns. ALS uses the prediction entropy for text generation. We proposed a concept which we call prominence, the margin to the strongest competing class, which measures how close a sample is to a class boundary in the feature space for adaptive smoothing. In Appendix C.4, we test whether the rules of FedLAS and ALS transfer to OSR, SmoothOP gives the best average OSCR on both splits for both objectives.

Appendix Table 9 (average OSCR over CUB, Aircraft and Cars; four post-processors; no Mixup; 3 seeds). Best in bold, second best underlined‾\underline{\text{underlined}}.

ObjectiveSplitLSALSFedLASOurs
SupCONEasy71.0471.97‾\underline{71.97}70.7572.03
SupCONHard65.2765.96‾\underline{65.96}65.7866.32
SpHOR-LiteEasy72.45‾\underline{72.45}71.1971.7472.91
SpHOR-LiteHard66.89‾\underline{66.89}65.7065.4067.18

(4) We proposed a rule for Mixup samples and validated it, which previous adaptive methods such as CaliMatch/FedLAS do not provide.

(5) This is not a standalone method, unlike CaliMatch, it integrates into four existing representation learning objectives.


Q1 (B) - Recent methods.

We added recent state-of-the-art training based OSR methods to Table 2: BackMix (TPAMI 2025), DEF (AAAI 2025), MEDAF (AAAI 2024) and NegSched/PosSched (ICLR 2026) as representative direct comparison methods. All use the same backbone, pretraining and splits as our experiments. On OSCR, our primary metric, the best result is a SmoothOP-augmented method in 5 of the 6 settings.

(1) Wang, Y., et al, BackMix: Regularizing Open Set Recognition by Removing Underlying Fore-Background Priors. TPAMI'25.

(2) Wang, Y., et al, Exploring diverse representations for open set recognition. AAAI'24.

(3) Kundu, A. K., et al, Boosting open set recognition performance through modulated representation learning. ICLR'26.

(4) Li, C., et al, Unlocking better closed-set alignment based on neural collapse for open-set recognition. AAAI'25.

(continued in Part 2/3)

AuthorsOpenReview ↗

**Official Response to Reviewer KLob (Part 2/3)**

CW 2(A) - Intra-class metric.

We thank the reviewer for this careful observation that led us to finding our error in the manuscript. Definition 4 contains a notation error: in our implementation, both metrics are computed on L2-normalized encoder features,

v^=Enc(x)∥Enc(x)∥2.\hat{v} = \frac{\mathrm{Enc}(x)}{\lVert \mathrm{Enc}(x) \rVert_2}.

Thus, Comp and Disp do not depend on the feature magnitude, and the results in Table 4 are not affected by changes in the feature norm, and directly supports our mechanism. We have corrected Definition 4 in the revised paper.


CW 2(B) - Per sample adjustment.

We respectfully disagree with the conclusion that it does not establish the contribution of the per-sample adjustment, and we maintain our position. The comparison treats the batch-mean control as if it were independent of our method, even though it is not. Both the batch mean and the per-sample adjustment are produced by prominence, and Table 11 shows that each of them contributes.

(1) Two components of the same signal. The per-sample coefficient decomposes exactly as εi=εˉ+(εi−εˉ)\varepsilon_i = \bar{\varepsilon} + (\varepsilon_i - \bar{\varepsilon}): a batch mean εˉ\bar{\varepsilon}, which changes over training, and a per-sample deviation, which assigns more or less smoothing within each batch. Both components are derived from prominence. The batch mean is the more effective component: of the mean OSCR gain over the baseline (0.62 pp on Easy, 0.54 pp on Hard), it accounts for 0.44 pp and 0.49 pp. The per-sample deviation adds a further gain on top of it (0.18 pp and 0.05 pp).

(2) The dominance of the batch mean does not invalidate the prominence explanation. The batch mean is not an external schedule: it is the general trend of prominence across all samples in the batch. As the training samples become prominent, the mean coefficient increases (App. D.2, Fig. 8). The batch-mean control therefore removes only the per-sample deviation; it keeps the contribution of prominence at the population level.

(3) The per-sample component is also important. The reversed-assignment control keeps the same batch mean and the same intra-batch variation, and only inverts which sample receives which coefficient. This degrades OSCR below the batch-mean control: inverting the assignment costs 0.25 pp (Easy) and 0.34 pp (Hard), compared with 0.18 pp and 0.05 pp for removing it. On Hard, inverting costs almost seven times more than removing. If the per-sample signal were noise, the two controls would perform alike. Δrev\Delta^{\text{rev}} is positive in 17 of 24 blocks, and the ordering Δrev>Δbm>0\Delta^{\text{rev}} > \Delta^{\text{bm}} > 0, stated as our hypothesis in App. D.1, holds on average on both splits.


Q2. Observer/Stage 2 Classifier.

Yes, with one difference, which we state explicitly.

(1) Equal for baselines and our augmented methods. All methods use the same backbone initialization, training schedule, seeds, and base smoothing coefficient (taken from Vaze et al., 2022). We clarified this by rewriting some of the paragraphs. We clarify that the weights of the linear probing classifier are initialized with that of the observer classifier.

(2) The one difference: selection of α. SmoothOP adds one hyperparameter, α\alpha. We select it on the training set and a surrogate dataset only: a three-fold cross-validation on a training split, with images from different fine-grained datasets as validation unknowns. We take the α\alpha that maximizes mean OSCR under MSP from the grid α∈{0,0.5,1,1.5,2,3,4,8}\alpha \in \{0, 0.5, 1, 1.5, 2, 3, 4, 8\}, which includes α=0\alpha = 0 (the baseline) (Lines 415-418). No test data is used for this selection.


Q3. Epsilon in Equations 7-8

We agree and we have updated the manuscript accordingly in Lines 379-381 and in equation (8) to state this claim more explicitly. We rechecked all our experiments to verify if they satisfy this condition and it does.

εi=2ε σ(−∣α∣ ui)∈(0, 2ε),ε≤12  ⇒  εi<1.\varepsilon_i = 2\varepsilon\,\sigma(-|\alpha|\,u_i) \in (0,\,2\varepsilon), \qquad \varepsilon \le \tfrac{1}{2} \;\Rightarrow\; \varepsilon_i < 1.

(continued in Part 3/3)

AuthorsOpenReview ↗

Official Response to Reviewer KLob (Part 3/3)

CW.3 + Q4: Notation mistakes

We corrected aforementioned inconsistencies, notation and descriptions and have marked them in blue in the corresponding locations in the revised manuscript.

(1) Mixup weight vs. smoothing coefficient (Line 378). λ\lambda now denotes only the Mixup weight (Eqs. 3, 10, 11); the base smoothing coefficient is denoted ε\varepsilon throughout (Eqs. 2, 8, 9).

(2) Encoder vs. projected features (Lines 237-238, 266-267). ff now denotes only the normalized projection, f(x)=Proj(Enc(x))/∥Proj(Enc(x))∥2f(x) = \mathrm{Proj}(\mathrm{Enc}(x))/\lVert \mathrm{Proj}(\mathrm{Enc}(x)) \rVert_2; the dd-dimensional encoder feature used in Stage 2 is denoted v=Enc(x)v = \mathrm{Enc}(x).

(3) Typos (Lines 89, 229). "from from" and "a a random permutation" are corrected.

In addition to these, we corrected many other inconsistencies, which are also marked in blue in the revised manuscript.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We thank all reviewers for their constructive feedback. We have uploaded a revised paper; all changes are marked in blue. The main revisions are:

(KLob, 4ohG) (1) Novelty and positioning. We clarified that SmoothOP is, to our knowledge, the first representation-based adaptive label smoothing for OSR, that it plugs into multiple different representation learning objectives, how it differs from classifier-based methods (CaliMatch/FedLAS), and why OSR needs a task-specific signal. Making label smoothing sample-adaptive is not by itself a recipe for success: when we apply other sample-adaptive LS variants such as FedLAS/ALS to OSR (App. C.4, Table 9), or reverse our own method's assignment (Table 11), the results are worse, while SmoothOP gives the best average OSCR on both splits for both objectives. Thus, the contribution goes beyond the per-sample form itself: it lies in finding a principled strategy for adaptive smoothing. In addition, to our knowledge, SmoothOP is the first adaptive label smoothing method to define a rule for Mixup samples.

(KLob, 4ohG) (2) Theoretical design principle. We contrasted the entropy-regularization view of classifier-based methods with our alignment view of label smoothing in representation learning, which we propose as a design principle. We then show how we build our method from the ground up. This gives SmoothOP a theoretical motivation, rather than empirical validation alone.

(KLob, BfBN) (3) Additional baselines. We added recent SOTA training-based OSR methods to Table 2 (ICLR'26, TPAMI'25, AAAI'25, AAAI'24). On OSCR, a SmoothOP-augmented method gives the best result in 5 of the 6 settings.

(BfBN, 4ohG) (4) Why fine-grained. We investigated in detail why SmoothOP is most effective in the challenging setting of fine-grained OSR and compared this with coarse-grained OSR (App. C.5). SmoothOP exploits a property of fine-grained datasets, such as high inter-class similarity, and thus cannot be expected to give the same gain on datasets without this property; on coarse-grained datasets, its performance neither degrades nor improves. Fine-grained recognition is a well-established field [1-8], and fine-grained OSR is its natural extension.

(KLob, BfBN, 4ohG) (5) Presentation. We corrected notation and other errors that we found, shortened the abstract and expanded the conclusion. We added analogies to make it easier to follow.

(KLob, 4ohG) (6) More requested ablations. We added the requested ablations, coarse-grained results, extra OpenAUC results, and other analysis to the Appendix and to our responses.

We thank Reviewer 4ohG for the engaged discussion and for acknowledging our empirical improvements. We are glad the discussion helped clarify most of the points from the initial review. We address each reviewer's points in the individual responses.

SmoothOP requires no change to the architecture or the objective, adds a single hyperparameter, and has negligible cost (Table 7), which makes it easy to adopt.

[1] Wei, X.-S., et al, Fine-Grained Image Analysis with Deep Learning: A Survey. TPAMI'22. [2] Lin, T.-Y., et al, Bilinear CNN Models for Fine-Grained Visual Recognition. ICCV'15. [3] Zheng, H., et al, Learning Multi-Attention Convolutional Neural Network for Fine-Grained Image Recognition. ICCV'17. [4] Yang, Z., et al, Learning to Navigate for Fine-grained Classification. ECCV'18. [5] Chen, Y., et al, Destruction and Construction Learning for Fine-grained Image Recognition. CVPR'19. [6] Zheng, H., et al, Learning Deep Bilinear Transformation for Fine-grained Image Representation. NeurIPS'19. [7] Zhuang, P., et al, Learning Attentive Pairwise Interaction for Fine-Grained Classification. AAAI'20. [8] Du, R., et al, Fine-Grained Visual Classification via Progressive Multi-Granularity Training of Jigsaw Patches. ECCV'20.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.