Refined Separate Class Learning (RSCL)

• RSCL is a method for long-tailed out-of-distribution detection that refines separate class learning by dynamically adjusting class-wise temperatures and mining informative outliers.
• It employs a shared backbone and a (C+1)-way classifier with dedicated head-specific and tail-specific branches to alleviate confusion between imbalanced ID classes and OOD samples.
• Evaluations on CIFAR10-LT, CIFAR100-LT, and ImageNet-LT show improved AUROC, AUPR, and balanced head-tail accuracy compared to previous approaches.

Searching arXiv for the RSCL paper and closely related long-tailed OOD detection work. Searching for arXiv paper ([2509.17034](/papers/2509.17034)) and related methods. Refined Separate Class Learning (RSCL) is a method for long-tailed out-of-distribution (OOD) detection that refines separate class learning (SCL) by introducing dynamic class-wise temperature adjustment and informative outlier mining. It is designed for settings in which in-distribution (ID) classes are imbalanced, with a few head classes having many samples and many tail classes having few samples, and in which OOD detection is degraded by confusion between OOD samples and both head and tail classes. RSCL uses a shared backbone feature extractor, a $(C+1)$-way classifier with an explicit outlier class, tail-specific learning, head-specific learning, and an affinity-based mining procedure for auxiliary OOD samples; empirically, it improves both OOD detection and ID classification on CIFAR10-LT, CIFAR100-LT, and ImageNet-LT (Feng et al., 21 Sep 2025).

1. Problem formulation in long-tailed OOD detection

OOD detection considers an ID training set composed of $C$ labeled classes and test inputs $x$ drawn from either the ID distribution $X_{\text{in}}$ or an unknown OOD distribution $X_{\text{out}}$. An OOD detector must assign ID inputs to their correct class and identify OOD inputs as outliers.

In the long-tailed setting, ID classes follow a long-tailed distribution: a few head classes have many samples, while many tail classes have few samples. This class imbalance skews decision boundaries and calibration. Tail-class samples are often confused with OOD samples because they are sparse and harder to represent well. OOD samples are often misclassified into head classes because head classes dominate the feature space and the classifier’s confidence. The skew also leads to poorly calibrated confidence: head classes tend to be overconfident, while tail classes tend to be underrepresented and uncertain.

These effects are central to why standard OOD detection degrades in long-tailed regimes. The problem is not only open-set rejection, but also the interaction between class imbalance, representation geometry, and confidence calibration. A plausible implication is that any effective method must separate OOD samples from head and tail classes in different ways rather than treating ID classes as a homogeneous set.

2. Position within separate class learning

Separate class learning (SCL) emerged as a solution that separately conducts head-specific and tail-specific class learning. Prior SCL methods identified in the RSCL formulation include PASCL, TSCL, and COCL. RSCL examines two limitations of existing works of SCL.

The first limitation is static temperature. Prior SCL methods use a single, fixed temperature across all pairs in contrastive or objective terms. OOD performance is sensitive to this choice, and a static temperature induces uniform pulling and pushing that does not align with class imbalance. The second limitation is uninformative outliers. Prior works often randomly sample auxiliary outliers, which can be uninformative and not aligned with head or tail confusion patterns, yielding limited gains.

RSCL addresses these limitations with two coordinated mechanisms. Dynamic class-wise temperature adjustment modulates the temperature parameter for each in-distribution class, and informative outlier mining identifies diverse types of outliers based on their affinity with head and tail classes. The training architecture consists of a shared backbone feature extractor and a $(C+1)$-way classifier, with two branches of learning: tail-specific learning (A-TSCL) and head-specific learning (A-OHL). During the first $3/4$ of training epochs, a mixed set of outliers from all three mined categories is used; in the final $1/4$, only “Neutral OOD” are used to refine discrimination.

3. Dynamic class-wise temperature adjustment

RSCL introduces a class-specific temperature that evolves across training epochs and depends on class frequency. For a sample $x$ of class $c(x) \in \{1,\ldots,C\}$, the dynamic class-wise temperature schedule is

$C$0

where $C$1 is a base temperature, $C$2 is the current epoch, $C$3 is the total number of epochs, and $C$4 is the $C$5-normalized class-frequency vector element constructed by normalizing $C$6 with $C$7, so that $C$8. This guarantees $C$9 and decreases with training and for head classes, since larger $x$0 makes $x$1 smaller (Feng et al., 21 Sep 2025).

The method uses temperature scaling inside contrastive objectives rather than in the final softmax for OOD scoring. Tail-specific A-TSCL uses $x$2 for interactions within tail classes. Head-specific A-OHL uses $x$3 for interactions between outliers and head samples. A global $x$4 is used for interactions involving the outlier prototype.

The stated intuition is twofold. In tail-specific learning, decreasing $x$5 within tail classes intensifies pulling among tail samples and tail prototypes, making tail clusters more compact and less confusable with OOD. In head-specific learning, decreasing $x$6 for head classes strengthens pushing between outliers and head samples, decreasing head overconfidence and separating OOD from head classes. Epoch dependence allows early training to be softer and late training to be sharper, while class dependence aligns the strength of interactions with class imbalance.

4. Informative outlier mining

RSCL distinguishes between uninformative outliers and informative outliers. Uninformative outliers are randomly sampled auxiliary samples with weak or irrelevant affinity to ID classes. Informative outliers are auxiliary samples selected based on quantified affinity to head or tail classes, enabling targeted learning of decision boundaries where confusion occurs.

The source of outliers is auxiliary OOD training data, with examples given as 300k Random Images for CIFAR-LT and ImageNet-Extra for ImageNet-LT. For an auxiliary sample $x$7, let $x$8 be the $x$9-way logits excluding the outlier class, and denote $X_{\text{in}}$0 as the number of head classes. The affinity-based outlier score is

$X_{\text{in}}$1

The paper describes this as “the sum of softmax probabilities over ID tail classes,” operationalized via the complement form using head classes (Feng et al., 21 Sep 2025).

The mining algorithm computes $X_{\text{in}}$2 for all auxiliary outliers and evenly splits them into three groups: “Tail-class-like OOD,” “Head-class-like OOD,” and “Neutral OOD.” The main text states that Algorithm 1 in Appendix B provides details, while precise thresholds, sampling ratios, and diversity strategies are not specified in the main text. Allocation is branch-specific: A-TSCL uses outliers to push against tail samples and prototypes, and A-OHL uses outliers to pull towards an outlier prototype and push away head samples. The first $X_{\text{in}}$3 of epochs use a mixed batch of all three outlier types to broaden OOD exposure, and the last $X_{\text{in}}$4 use only Neutral OOD to refine discrimination on ambiguous cases.

5. Objective functions and decision rule

RSCL extends the classifier to $X_{\text{in}}$5 classes, where the extra class is the outlier class. The outlier class learning objective is

$X_{\text{in}}$6

where $X_{\text{in}}$7 is cross-entropy, $X_{\text{in}}$8 are logits, and $X_{\text{in}}$9 is a weight for the outlier class term.

For tail-specific class learning (A-TSCL), let $X_{\text{out}}$0 be a projection head output of the penultimate feature. A batch $X_{\text{out}}$1 of ID samples is split into head $X_{\text{out}}$2 and tail $X_{\text{out}}$3; $X_{\text{out}}$4 is an anchor; $X_{\text{out}}$5 are positives from the same tail class; and $X_{\text{out}}$6 are informative outliers. Tail prototypes $X_{\text{out}}$7 are derived from classifier weights of tail classes using an MLP and $X_{\text{out}}$8-normalization. The per-anchor contrastive loss is

$X_{\text{out}}$9

The tail branch averages over $(C+1)$0 in $(C+1)$1, with $(C+1)$2:

$(C+1)$3

For head-specific class learning (A-OHL), $(C+1)$4 is an outlier anchor, and the outlier-class prototype $(C+1)$5 is derived by taking the $(C+1)$6-th classifier weight, passing it through an MLP, and $(C+1)$7-normalizing it. The per-outlier contrastive loss is

$(C+1)$8

The head branch average is

$(C+1)$9

The full RSCL objective is

$3/4$0

where $3/4$1 and $3/4$2 weight the tail and head branches. The formulation explicitly notes that RSCL does not introduce energy-based terms in the objective and does not use OE-style confidence penalties in the main formulation (Feng et al., 21 Sep 2025).

At test time, the OOD score uses the $3/4$3-th softmax probability:

$3/4$4

where $3/4$5 is the softmax score of the outlier class and $3/4$6 is a threshold. The threshold is chosen to satisfy a desired detection operating point, such as FPR@95 on OOD, following prior long-tailed OOD works including PASCL, EAT, and TSCL.

6. Experimental profile and reported results

The reported ID datasets are CIFAR10-LT and CIFAR100-LT with imbalance ratio $3/4$7, and ImageNet-LT. OOD training uses 300k Random Images for CIFAR-LT and ImageNet-Extra for ImageNet-LT. OOD test uses six datasets—Texture, SVHN, CIFAR, Tiny ImageNet, LSUN, and Places365—for CIFAR-LT, and ImageNet-1k-OOD for ImageNet-LT. The evaluation metrics are AUROC, AUPR, FPR@95, and ID accuracy (ACC).

On CIFAR10-LT with ResNet-18, averaged over six OOD test sets, RSCL vs PASCL is reported as AUROC 93.46 vs 90.53 (+2.93), AUPR 93.28 vs 88.49 (+4.79), FPR95 27.73 vs 32.99 (−5.26), and ACC 79.14 vs 75.94. On CIFAR100-LT with ResNet-18, averaged over six OOD test sets, RSCL vs PASCL is reported as AUROC 76.11 vs 73.10 (+3.01), AUPR 70.76 vs 67.06 (+3.70), FPR95 62.86 vs 67.70 (−4.84), and ACC 42.81 vs 39.95. On ImageNet-LT with ResNet-50 and ImageNet-1k-OOD, RSCL reports AUROC 78.90, AUPR 78.10, FPR95 73.04, and ACC 46.19, and is described as consistently higher in OOD detection and ID accuracy than PASCL, EAT, COCL, and TSCL (Feng et al., 21 Sep 2025).

The head-versus-tail accuracy breakdown indicates that RSCL improves both head and tail classes. For CIFAR10-LT, RSCL head is 87.53, tail is 70.44, and overall is 79.14. For CIFAR100-LT, RSCL head is 62.76, tail is 22.86, and overall is 42.81. For ImageNet-LT, RSCL head is 61.57, tail is 30.66, and overall is 46.12. The reported interpretation is that RSCL strikes a better balance between head and tail performance.

Ablation on CIFAR100-LT, averaged across six OOD sets, separates the effects of informative outlier mining and dynamic temperatures. Adding informative outlier mining to LOCL plus vanilla tail and head losses improves AUROC from 74.15 to 75.29, FPR95 from 65.81 to 63.33, and ACC from 40.11 to 41.70. Adding dynamic temperatures improves AUROC from 74.15 to 75.23, FPR95 from 65.81 to 64.88, and ACC from 40.11 to 42.05. Full RSCL, described as LOCL + tail + head + OM + TA, achieves AUROC 76.11, AUPR 70.76, FPR95 62.86, and ACC 42.81. The qualitative analysis based on t-SNE plots states that RSCL produces more separated clusters for tail classes and pushes OOD away from head and tail regions compared to training with LOCL alone.

7. Implementation considerations, assumptions, and limitations

The high-level training pipeline begins with $3/4$8 (long-tailed) and $3/4$9 (auxiliary OOD), a shared feature extractor, a $1/4$0-way classifier, projection heads for $1/4$1, and MLPs to form tail prototypes and an outlier prototype. Outlier mining computes $1/4$2 for each $1/4$3 and evenly splits $1/4$4 into Tail-like OOD, Head-like OOD, and Neutral OOD. For epoch $1/4$5, the method computes $1/4$6, constructs mini-batches from $1/4$7 and $1/4$8, uses mixed OOD categories if $1/4$9 and only Neutral OOD otherwise, computes $x$0, $x$1, and $x$2, and updates parameters by minimizing $x$3. At inference, it thresholds $x$4 to detect OOD and otherwise classifies with $x$5 over the $x$6 ID softmax components.

The main hyperparameters identified are the base temperature $x$7, epoch schedule $x$8, weights $x$9, $c(x) \in \{1,\ldots,C\}$0, and $c(x) \in \{1,\ldots,C\}$1, the mining thresholds to split $c(x) \in \{1,\ldots,C\}$2 into three categories, the mixing schedule ($c(x) \in \{1,\ldots,C\}$3 versus $c(x) \in \{1,\ldots,C\}$4), and the batch composition for $c(x) \in \{1,\ldots,C\}$5 and $c(x) \in \{1,\ldots,C\}$6. The main text does not report explicit numeric settings and refers to Appendix D.2 for implementation details. Dynamic temperature has negligible overhead because it requires a per-class scalar update each epoch. Outlier mining requires a forward pass to compute $c(x) \in \{1,\ldots,C\}$7 over the auxiliary set; the stated memory and runtime overhead is modest relative to standard OE pipelines.

The limitations and assumptions are explicit. Availability and quality of auxiliary outliers matter, and poor auxiliary data limits gains. Sensitivity to the temperature schedule and mining thresholds may affect robustness, while the paper does not provide detailed hyperparameter recipes in the main text. Severe domain shifts not represented in auxiliary data may remain challenging. The paper uses standard backbones—ResNet-18 for CIFAR-LT and ResNet-50 for ImageNet-LT—and states that, for other modalities or datasets, one can reuse the dynamic temperature schedule and category-aware outlier mining while adjusting the $c(x) \in \{1,\ldots,C\}$8-way classifier accordingly (Feng et al., 21 Sep 2025).