WorksComputer Science
When Neural Collapse Fails to Route: Routing Drift in Continual Learning
Abstract
Progressive neural-collapse classifiers can learn strong task-local decision boundaries while failing the global Class-IL decision. We identify the failure as routing drift: under an expanding equiangular tight frame (ETF) classifier, a held-out sample can remain separable within its true task block, yet receive its largest global score from another block. This produces high Task-IL accuracy but low Class-IL accuracy, and small replay buffers can mask the problem because stored samples are optimized repeatedly and often remain well routed. We formalize route accuracy directly from the global class argmax and use it to separate within-task discrimination from cross-task score comparability. This decomposition suggests a classifier-level correction: preserve the global ETF frame, but add class-local reliability evidence when transported class measures deviate from the neural-collapse ideal. We develop Reliability-Calibrated Routing (RCR), a lightweight post-hoc classifier that composes standardized ETF alignment with radial plausibility around transported, NC-shrunk class representatives. Across three datasets and two fixed replay-buffer budgets, this NC-consistent correction gives the best Class-IL FAA on five of six settings and reduces Class-IL forgetting on all six. The largest gains occur in harder routing regimes: up to FAA over the reported reference, and up to FAA and final route accuracy in paired evaluations, while preserving Task-IL performance. These results position routing drift as a measurable failure mode of ETF-based continual learning and show that reliability-calibrated class evidence can repair part of the lost global routing without retraining the representation. Code and evaluation artifacts are available at https://towzeur.github.io/RCR/.
Introduction
Class-incremental learning (Class-IL) asks a model to classify among all classes observed so far, without receiving the task identity at test time. Task-incremental learning (Task-IL) is easier: the correct task block is known, so the classifier only chooses among classes inside that block. The gap between Class-IL and Task-IL therefore measures more than forgetting; it reveals whether evidence remains comparable across task blocks [3, 25, 28]. Prior analyses decompose Class-IL into selecting the correct task and selecting the correct class within that task [16, 17, 22]. We study the score-level form of this decomposition inside progressive ETF classifiers.
Neural-collapse-inspired continual learners use equiangular tight frame (ETF) classifiers to impose a structured global geometry on class prototypes [27]. ProNC is a representative example: it expands the ETF as new classes arrive and learns features aligned to the evolving frame [36]. This geometry yields strong Task-IL performance, which means task-local class separation is often preserved. Yet our diagnostics show that many remaining Class-IL errors are not same-task confusions. The correct class can be identifiable once the true task block is known, while the global ETF score assigns the sample to another block.
Figure 1 illustrates this score-level failure. The true class has the largest logit inside its task block, so the Task-IL decision is correct, but another block contains the global maximum. We call this routing drift. In ETF continual-learning runs, Task-IL can remain high while route accuracy collapses, indicating that the model keeps task-local discrimination but loses global route comparability.

Figure 1. Routing drift as a logit-level failure. The correct class can win within its task block while a different block contains the global argmax. Routing drift is therefore not a within-task class error; it is a cross-task score-comparability error.
This paper studies routing drift as a geometric failure mode of neural-collapse continual learning. The point is not that ETF classifiers fail to separate classes locally; rather, their global scores become unreliable across task blocks. We therefore ask whether Class-IL can be improved by correcting cross-task competition while preserving the task-local structure already learned by the progressive ETF model. Our answer is Reliability-Calibrated Routing (RCR), a reliability-compositional score that combines global ETF alignment with radial plausibility around transported class measures. RCR targets the inference geometry of progressive ETF classifiers: it keeps the global neural-collapse frame, adds local class reliability, and does not use an external task predictor. It is a post-hoc classifier-level correction, not a new representation learner; this scope is deliberate because it isolates the routing failure from the underlying training dynamics. The method is designed for progressive ETF classifiers, and extending the same reliability principle beyond explicit ETF frames is left open.
Our contribution is threefold. (i) We define route accuracy as a diagnostic read directly from the global ETF prediction, specializing the known Class-IL decomposition to progressive neural-collapse classifiers. (ii) We show that replay buffers can mask routing drift, so replay-fitted task calibration is not a reliable substitute for held-out routing evaluation. (iii) We introduce RCR, a reliability-calibrated classifier that preserves the neural-collapse frame, models local class plausibility, gives the best Class-IL FAA on five of six dataset-buffer settings, and reduces Class-IL forgetting on all six while preserving Task-IL accuracy.
Related Work
Class-incremental continual learning. Class-incremental learning evaluates a model after each task over all classes seen so far, without providing the task identity at test time [24]. Early and still influential approaches regularize important parameters or logits [18, 21, 41]. or rehearse previous data through direct replay [29, 30], exemplar classifiers [28], gradient-aware replay [1, 5, 23], and regularized replay [2, 3]. Other approaches use cosine classifiers, pooled-output distillation, feature boosting, prompts, contrastive objectives, coreset selection, provable memory construction, and stability-inducing weight perturbations [4, 6, 7, 8, 15, 31, 33, 34, 37]. These methods address forgetting in the representation and in the classifier, but Class-IL also requires scores from different tasks to remain comparable. Prior analyses separate this requirement into two events: selecting the correct task block and selecting the correct class inside that block [16, 17, 22]. We use this decomposition diagnostically, but do not train a task predictor. Route accuracy is read directly from the global class argmax, and RCR changes the class score itself rather than fitting a separate task-id module.
Neural collapse and ETF classifiers. Neural collapse describes a terminal training geometry in which within-class variability shrinks, class means align with classifier weights, and these directions form a simplex ETF [27]. Subsequent work has analyzed its dynamics, layer-peeled models, imbalanced regimes, and feature-classifier alignment beyond standard balanced classification [9, 13, 38, 39]. This geometry has motivated fixed ETF classifiers for continual learning, where the classifier is prescribed rather than learned from scratch at every task [40]. Neural-collapse-inspired contrastive continual learning further uses predefined hard prototypes and simplex-structure distillation to reduce representational overlap and stabilize learning across tasks [26]. Progressive neural-collapse methods avoid committing to a final class set by expanding the ETF as new classes arrive, then using nearest-ETF inference in the current frame [36]. This design gives a simple global geometry for all classes seen so far. Our work starts from this trained progressive ETF model and asks a different question: after several feature-frame updates, is the nearest-ETF score still calibrated for both old and new classes? The answer motivates a classifier-level correction rather than a new ETF construction.
Prototype drift and local reliability. Prototype and cosine classifiers are widely used when class evidence is expressed by normalized features, class means, or learned weight vectors [28, 35]. They are also sensitive to representation drift: an old class mean estimated in one feature frame may no longer be reliable after the backbone changes. Learnable Drift Compensation (LDC) estimates a previous-to-current feature projector and applies it to old class prototypes before nearest-class-mean inference [10]. This line of work shows that prototype positions should be interpreted in the current feature frame, not as fixed historical measurements. A complementary issue is metric reliability. FeCAM models class distributions with covariance-aware distances, showing that a single isotropic nearest-prototype rule can be too restrictive when local class geometry differs across classes [11]. We use these observations in a constrained way: transported prototypes and competitive radii provide local reliability evidence, while the explicit ETF classifier remains the global neural-collapse frame. Thus RCR is not transported NCM, not a replay-fitted calibration head, and not a replacement for ProNC’s progressive ETF.
Routing terminology. In many-task learning, task routing can denote an architectural mechanism that selects task-specific paths [32]; recent mixture-of-experts continual-learning papers similarly study expert assignment drift when samples are reassigned to newly added experts [12, 42]. These works use an explicit router, expert assignment, or task-specific execution path. Our setting is different: there is one shared classifier, no test-time task identity, and no learned routing head. Routing drift denotes inter-block score incomparability inside the same global classifier. This distinction is important experimentally: improving route accuracy here means changing the class evidence used by the global argmax, not adding a task oracle in front of the classifier.
Preliminaries
Continual learning protocol. We consider a sequence of tasks . Task provides data whose labels are disjoint from previous tasks. Let be the set of classes observed after task , and let map class to its task block. In Class-IL, the model predicts among all classes in without receiving task identity. In Task-IL, prediction is restricted to the true task block. We report final average accuracy (FAA) and final forgetting (FF). If is the accuracy on task after training task , then
Primary notation is summarized in the supplementary material (Table S2).
| Type | Symbol | Description |
| Protocol and indexing | ||
| Set | Ordered continual-learning task sequence. | |
| Index | Current training or evaluation stage. | |
| Set | Training data introduced by task . | |
| Set | Classes observed after task . | |
| Scalar | Number of classes per task block. | |
| Function | Task block containing class . | |
| Set | Replay samples stored for class . | |
| Metric | FAA, FF | Final average accuracy and final forgetting. |
| Features, vectors, and matrices | ||
| Function | Current feature extractor. | |
| Vector | Normalized current feature, . | |
| Vector | Normalized feature from the previous model. | |
| Vector | Progressive ETF vertex for class . | |
| Matrix | Ridge previous-to-current feature transport at stage . | |
| Vector | Transported class representative in the current feature frame. | |
| Vector | NC-shrunk class representative used by the radial score. | |
| Operator | Cosine-distance convention for normalized vectors. | |
| Operator | Per-sample standardization over currently seen classes. | |
| Scores, radii, and scalars | ||
| Score | ETF cosine score for feature and vertex . | |
| Score | Radius-normalized radial score around the NC-shrunk class representative. | |
| Score | Final RCR score combining standardized ETF and radial evidence. | |
| Scalar | Raw class radius estimated from replay features. | |
| Scalar | Mean raw radius over currently seen classes. | |
| Scalar | Competitive class-dependent reliability radius. | |
| Scalar | Mean competitive radius. | |
| Scalar | Radius-compression exponent used to form competitive radii. | |
| Scalar | NC-shrink coefficient used to retract toward . | |
| Scalar | Exponent controlling the shrink coefficient. | |
| Scalar | Fixed ETF/radial mixture weight, . | |
| Scalar | Numerical stability constant. | |
| Losses and routing diagnostics | ||
| Loss | Reference ProNC training objective. | |
| Loss | Supervised cross-entropy term. | |
| Loss | ETF-alignment term in the reference objective. | |
| Loss | Feature-distillation term from the previous model. | |
| Score | Maximum class score inside task block . | |
| Scalar | True-block score margin, . | |
| Event | Indicator that the global argmax lies in the correct task block. | |
| Event | Indicator that the correct class wins inside the true task block. | |
| Metric | Route accuracy | Test-set expectation of . |
Table S2. Primary notation. The first column makes each object’s type explicit. We use calligraphic letters for sets, lower-case symbols for feature/prototype vectors and scalars, upper-case Roman letters for matrices, and for losses.
Route accuracy. Following the decomposition discussed in § 2, we define a route diagnostic directly from the global class prediction. For a sample and a global prediction ,
Route accuracy is the expectation of over the test set. It is not an oracle available to the method, but a measurement of whether the class argmax falls inside the correct task block. High Task-IL but low route accuracy means that features remain separable within the correct block, but the global classifier assigns them to another block.
Neural collapse and ETF inference. Neural collapse describes the terminal regime in which centered class means and classifier weights align to a simplex equiangular tight frame (ETF) [27]. We recall the standard NC1–NC4 definitions in the supplementary material (Section C). A -class simplex ETF is a set of unit vectors satisfying
Under the neural-collapse idealization, classification reduces to nearest-prototype inference in the ETF frame:
Progressive Neural Collapse. ProNC progressively constructs the ETF target during continual learning rather than fixing a global ETF in advance [36]. After task 1, it extracts an ETF aligned with the learned first-task class means. Before each later task, it expands the ETF by adding vertices for new classes while keeping the old geometry close to the previous frame. For task , the reference training objective combines supervised learning, alignment to the expanded ETF, and feature distillation from the previous model:
where is the normalized feature from the previous model. At evaluation, the linear classifier is replaced by nearest-ETF inference:
This nearest-ETF rule is the reference classifier analyzed below. Definitions of and are given in the supplementary material (Section B).
The Routing Drift Problem
ProNC gives every seen class an ETF vertex and compares normalized features to all vertices in one shared frame. This is the right geometry for global Class-IL: all class vertices have equal norm and equal pairwise angle, so the classifier layout itself does not favor a task block. That same symmetry makes the failure interpretable. If a wrong task block wins under an ETF head, the problem is not an uneven classifier layout. It is a mismatch between the symmetric decision rule and the class feature distributions that occupy it. Such a mismatch can emerge after several tasks because old classes are represented through replay and distillation, whereas new classes are observed densely in the current task.
To make the failure explicit, fix an evaluation stage and let index task blocks. Write the best score inside block as
As summarized in Figure 2, Class-IL succeeds only if the true block wins the cross-task competition and the correct class wins inside that block. Task-IL measures only the second event. Hence a model can have high Task-IL while failing globally whenever

Figure 2. Class-IL needs both events. A prediction is globally correct only when the true task block wins and the correct class wins inside that block.
We call this routing drift: features remain discriminative within their own task block, but their global ETF scores no longer keep the correct block ahead. In neural-collapse terms, this is a violation of equal class-measure reliability, not a failure of the ETF geometry itself.
This gives a direct link between routing and Class-IL. Define the block-restricted prediction
Let be the routing event, and let be the task-local event obtained by restricting prediction to the true block. Exact Class-IL correctness requires both events:
Thus
Task-IL estimates the block-local event , whereas Class-IL also requires the route . When task-local discrimination is preserved, improving route accuracy directly improves achievable Class-IL. This factorization is consistent with prior analyses that decompose Class-IL into within-task prediction and task-identity prediction [16, 17, 22]. Here, route accuracy is not produced by a learned task-id predictor and is not an oracle; it is read from the global ETF classifier’s own argmax.
An example block-score profile is shown in the supplementary material (Figure S3); the main diagnostic in Figure 3 shows that this is not an isolated sample-level artifact but a sequence-level failure mode.
Figure 3. Held-out errors are dominated by wrong-task routing. On Seq-CIFAR-100 with buffer 200, Task-IL remains high across tasks, but Class-IL follows route accuracy downward. The right panel decomposes global predictions: late errors are mostly wrong-task routing rather than same-task class confusion.
Figure S3. A routing error at the block-score level. For one held-out Seq-CIFAR-100 sample, the true class wins within its own task block, but the ProNC ETF maximum is larger in a wrong block. RCR changes the cross-block score profile so the correct block wins, without using task identity at test time.
Replay is also an unreliable route-calibration set. Stored samples are optimized repeatedly and can remain close to their vertices: in Figure 4, replay routing is almost perfect at the final task, while held-out test samples have only 42.81% route accuracy and 57.19% wrong-task error. A useful correction should therefore preserve the ETF frame, avoid fitting a task oracle from the buffer, and add local class reliability directly to the global scoring rule. Replay labels reveal which task a stored sample came from, but they do not measure how held-out old samples compete against newer task blocks. The classifier must instead keep the global ETF comparison intact and adjust class evidence only when the measured class
Figure 4. Replay routing is optimistic. Replay samples stay well routed while held-out test routing collapses on Seq-CIFAR-100/200; mean std over three seeds.
distribution has drifted away from its ideal NC geometry. This is the motivation for RCR, which we introduce in the next section.
Method: Reliability-Calibrated Routing
From Routing Drift to Reliability-Calibrated Scores
Routing drift calls for a classifier-level correction. The global ETF frame should remain the shared comparison space, but each class score should also reflect how reliable its local class measure is in the current feature frame. Let and let be the ETF vertex for class . The reference score is
This score is globally comparable when all classes follow the same neural-collapse geometry, but it treats every vertex as an equally reliable point representative. RCR keeps this global NC/ETF evidence and adds local reliability around a transported prototype and radius .
Transported Class Measures
Local reliability must be measured after feature drift. Following the prototype-drift motivation of LDC [10], we estimate a previous-to-current linear transport. Unlike transported NCM, however, we use transported prototypes only as local anchors; the ETF vertices remain the global classifier. Thus is not a replacement for . It is the measured location of class in the current feature frame. Figure 5 illustrates this separation.
Figure 5. RCR changes the route boundary without changing the backbone. A held-out CIFAR-100 feature is shown in a local two-dimensional slice passing through the original high-dimensional feature, so has the same scores as in evaluation. The black curve is the route boundary between the correct task block and the ETF-predicted wrong block in that slice. (a) Under ETF scoring, lies in the red region: a wrong task block wins globally even though nearby features from the correct class remain coherent. (b) RCR keeps ETF evidence but adds transported radial plausibility around , moving the route boundary so the same feature falls in the correct-task region. Filled stars denote current ETF vertices, hollow stars denote previous prototypes, and triangles denote transported prototypes. Classes are selected only to visualize one corrected routing case; this selection is not used for training, hyperparameter choice, or evaluation.
At the end of task , the previous frozen backbone and the current backbone provide paired features on the available training stream. We fit the same form of ridge linear transport for all classes,
For old classes, the previous prototype is moved into the current frame by
For new classes, is the normalized centroid of current-backbone features. All class representatives now live in the same frame, but they are not used as classifier weights.
Radial Reliability
The transported prototype gives a class location, not its reliability. Under exact NC1, all classes are equally concentrated. In continual learning, replay and drift make this concentration class-dependent. We measure this local NC1 violation while keeping as the global NC2/NC3 reference. Let denote the cosine-distance convention used by the radial head; for normalized vectors this is their squared Euclidean distance, and the implementation only clamps it at zero for numerical stability. Let be the replay samples of class available after task . We estimate the raw class radius by
If a class is absent from replay, we use the global mean radius as a fallback.
The raw radii cannot be unconstrained class scales. Since the ETF allocates class directions uniformly, no class should win by claiming an arbitrarily large region of the sphere. We therefore convert raw radii into a competitive radius:
where is the mean raw radius. This keeps the average radial budget fixed while allowing measured dispersion to modulate local support; the exponent makes the modulation conservative. Let denote the resulting mean competitive radius. This normalization is not a softmax: increasing one class radius increases its local support only relative to a fixed total radius budget, so the other classes must compensate through the denominator.
Finally, using directly would let the local class measure replace the NC reference. As a stabilization step, we retract it toward according to relative reliability:
with and . Classes with tight support keep more of their transported center, while high-radius classes are pulled back toward the symmetric ETF vertex. The upper clip prevents full trust in a drift-compensated prototype; the unconstrained variant is reported in the supplement. The radial score is the negative radius-normalized cosine distance to this NC-retracted center:
A class scores highly only when the query is close to its NC-retracted prototype relative to its estimated radius. This term is class-local: it does not predict a task block or introduce a separate router.
Final Classifier
ETF scores and radial scores have different units, so we standardize each score vector per sample. This removes sample-specific scale and offset before composition; the final decision depends on the relative class ordering supplied by each evidence source, not on the arbitrary numerical units of cosine alignment versus radius-normalized distance:
The final RCR score is
where is the number of classes per task. The scalar is fixed by task granularity and shared by all classes and samples. It is not selected from validation data: it increases the radial contribution when each task block contains more competing classes, and keeps more ETF weight in small-block settings such as Seq-CIFAR-10.
RCR does not retrain the backbone, fit a task oracle, use test-time task identity, or add a task-age prior. The only class-specific quantities are the ETF vertex , transported prototype , and reliability radius . No scalar is selected per dataset: all reported datasets use the same fixed rules , , , and .
NC consistency. RCR is designed to be inactive at the exact neural-collapse fixed point. Assume that transported class representatives coincide with ETF vertices, , and that all classes have the same radius, . Then by Equation (18) for any clipped value of , and
The radial score is therefore a positive affine transform of the ETF score. Because Equation (21) uses per-sample standardization, , and RCR reduces exactly to the normalized nearest-ETF classifier. Thus, RCR changes decisions only when progressive neural collapse is imperfect, i.e. when transported class centers deviate from ETF vertices or class reliabilities become heterogeneous.
Experiments
Experimental Setup
Benchmarks and evaluation setting. We evaluate on Seq-CIFAR-10, Seq-CIFAR-100 [19], and Seq-TinyImageNet [20] with replay buffers of 200 and 500 samples. Seq-CIFAR-10 is split into 5 tasks with 2 classes per task, while Seq-CIFAR-100 and Seq-TinyImageNet use 10 tasks with 10 and 20 classes per task, respectively. To isolate classifier geometry, all paired head comparisons use the same learned progressive-ETF representations and differ only in the evidence used for global prediction. Reproduced artifacts follow the same ResNet-style backbone family as the reference implementation [14]. Additional implementation details, including the ProNC training objective and the values of used to produce the checkpoints, are given in the supplementary material (Section B).
Baselines and paired evaluation. For the literature tables, all non-Ours baselines are taken from [36], which reports the same datasets, memory budgets, and FAA/FF metrics. For the paired head comparison in Table 1, we reproduced the reference model with the official code over seeds 0, 1, 2 and evaluated the ETF head and ProNC + RCR on the same artifacts. The reproduced ETF head is close to, but not identical to, the published row: across the six dataset–buffer settings, the absolute difference in final Class-IL FAA ranges from 0.29 to 2.45 points. We use published rows for literature positioning and reproduced artifacts for paired classifier comparisons.
| Dataset | Buffer | Score | Class-IL | Task-IL | Route | ||
| FAA ↑ | FF ↓ | FAA ↑ | FF ↓ | ↑ | |||
| CIFAR-10 | 200 | sETF | 66.14 | 31.47 | 96.74 | 0.62 | 66.58 |
| CIFAR-10 | 200 | sRCR (Ours) | 70.36 (+4.22) | 22.85 (-8.62) | 96.67 (-0.07) | 0.73 (+0.11) | 70.82 (+4.24) |
| CIFAR-10 | 500 | sETF | 72.16 | 24.18 | 96.82 | 0.47 | 72.71 |
| CIFAR-10 | 500 | sRCR (Ours) | 73.68 (+1.52) | 17.73 (-6.45) | 96.81 (-0.01) | 0.54 (+0.07) | 74.33 (+1.62) |
| CIFAR-100 | 200 | sETF | 41.71 | 38.55 | 85.53 | 4.51 | 42.81 |
| CIFAR-100 | 200 | sRCR (Ours) | 49.74 (+8.03) | 16.27 (-22.28) | 87.28 (+1.75) | 2.36 (-2.15) | 51.43 (+8.62) |
| CIFAR-100 | 500 | sETF | 48.65 | 23.16 | 85.54 | 4.29 | 50.14 |
| CIFAR-100 | 500 | sRCR (Ours) | 53.27 (+4.62) | 10.18 (-12.98) | 87.74 (+2.20) | 2.02 (-2.27) | 55.04 (+4.90) |
| TinyImageNet | 200 | sETF | 25.77 | 43.45 | 68.30 | 9.58 | 28.58 |
| TinyImageNet | 200 | sRCR (Ours) | 31.42 (+5.65) | 23.84 (-19.61) | 69.59 (+1.29) | 8.61 (-0.97) | 34.76 (+6.18) |
| TinyImageNet | 500 | sETF | 26.61 | 41.16 | 67.71 | 9.78 | 29.37 |
| TinyImageNet | 500 | sRCR (Ours) | 32.24 (+5.63) | 22.47 (-18.69) | 70.16 (+2.45) | 7.90 (-1.88) | 35.46 (+6.09) |
Table 1. RCR improves the global decision by repairing route accuracy. On the same frozen ProNC checkpoints, Class-IL and Route increase together while Task-IL remains nearly unchanged. Parentheses report deltas against sETF; all entries are averaged over seeds 0, 1, 2.
Metrics. We report final average accuracy (FAA) and final forgetting (FF) in both Class-IL and Task-IL. Class-IL evaluates the global classifier over all seen classes. Task-IL restricts prediction to the ground-truth task block and therefore measures task-local class discrimination. To quantify the failure mode directly, we also report route accuracy: the fraction of test samples for which the globally predicted class belongs to the correct task block. A routing correction should increase route accuracy and Class-IL while leaving Task-IL nearly unchanged.
Main Results
Paired classifier comparison
Table 1 compares the reference ETF score and ProNC + RCR on identical evaluation artifacts. ProNC + RCR improves Class-IL FAA and Route together, while Task-IL changes little. This is the signature predicted by Equation (11): the gain comes from selecting the correct task block, not from relearning within-task boundaries.
Across datasets and memory budgets, Class-IL gains track Route gains. After progressive ETF training, the model often still separates classes within the correct task block, but the global score assigns samples to competing blocks. A compact visual summary and task-by-task diagnostic curves are provided in the supplementary material (Figure S1 and Figure S2).
Figure S1. Task-by-task diagnostic curves across all validated datasets. Each row shows one dataset and memory budget; columns report Class-IL accuracy, Task-IL accuracy, and route accuracy over the continual sequence. The same frozen checkpoints are evaluated with the ProNC ETF head and ProNC + RCR. Across datasets, RCR primarily changes the Class-IL and Route curves while leaving Task-IL close to the reference head, supporting the routing-drift interpretation.
Figure S2. Final head comparison across datasets and memory budgets. Each panel compares the ProNC ETF head with ProNC +RCR on the same saved evaluation artifacts. Rows correspond to buffer sizes 200 and 500; columns report Class-IL FAA, Class-IL FF, Task-IL FAA, Task-IL FF, and final route accuracy.
Comparison against the state of the art
The two tables below are for literature positioning. Their prior-method rows are copied from the published reference table, while the Ours row is computed on our three-seed reproduction. The perfectly paired comparison is therefore Table 1, where both heads are evaluated on the same reproduced checkpoints; the literature tables report whether the resulting classifier correction reaches or exceeds the published numbers under the same dataset and memory budgets.
Table 2 and Table 3 report the main comparison for replay buffers of 200 and 500 samples. ProNC + RCR sets the best Class-IL FAA on five of six dataset-buffer settings and reduces Class-IL FF on all six relative to the reported reference. These tables should be read as the effect of a post-hoc, NC-consistent classifier correction applied to a trained progressive ETF model, not as a replacement training algorithm for continual learning. The largest gains appear on harder routing regimes, where many task blocks compete globally.
| Setting | Method | Venue | Seq-CIFAR-10 FAA ↑ | Seq-CIFAR-10 FF ↓ | Seq-CIFAR-100 FAA ↑ | Seq-CIFAR-100 FF ↓ | Seq-TinyImageNet FAA ↑ | Seq-TinyImageNet FF ↓ |
| Class-IL | ER [29] | ICLR 2019 | 44.79 ± 1.86 | 59.30 ± 2.48 | 21.78 ± 0.48 | 75.06 ± 0.63 | 8.49 ± 0.16 | 76.53 ± 0.51 |
| Class-IL | iCaRL [28] | CVPR 2017 | 49.02 ± 3.20 | 23.52 ± 1.27 | 28.00 ± 0.91 | 47.20 ± 1.23 | 7.53 ± 0.79 | 31.06 ± 1.91 |
| Class-IL | DER [3] | NeurIPS 2020 | 61.93 ± 1.79 | 35.79 ± 2.59 | 31.23 ± 1.38 | 62.72 ± 2.69 | 11.87 ± 0.78 | 64.83 ± 1.48 |
| Class-IL | DER++ [3] | NeurIPS 2020 | 64.88 ± 1.17 | 32.59 ± 2.32 | 28.13 ± 0.51 | 60.99 ± 1.52 | 11.34 ± 1.17 | 73.47 ± 1.23 |
| Class-IL | Co2L [4] | ICCV 2021 | 51.27 ± 1.86 | 30.17 ± 8.57 | 18.09 ± 0.49 | 64.14 ± 0.69 | 12.95 ± 0.06 | 62.04 ± 0.28 |
| Class-IL | CILA [37] | ICML 2024 | 59.68 ± 0.65 | 37.52 ± 5.84 | 19.49 ± 0.53 | 64.01 ± 0.18 | 12.98 ± 0.10 | 63.11 ± 0.61 |
| Class-IL | MNC3L [6] | WACV 2025 | 52.20 ± 1.56 | 33.74 ± 1.65 | 15.81 ± 0.48 | 62.51 ± 0.40 | 10.57 ± 1.66 | 59.68 ± 2.02 |
| Class-IL | STAR [8] | ICLR 2025 | 62.10 ± 2.21 | 21.78 ± 3.16 | 18.29 ± 2.58 | 68.40 ± 8.52 | 11.55 ± 3.23 | 67.43 ± 1.48 |
| Class-IL | CSReL [33] | ICLR 2025 | 37.46 ± 1.57 | 26.34 ± 2.18 | 29.06 ± 1.00 | 58.23 ± 1.54 | 18.14 ± 3.10 | 49.77 ± 2.27 |
| Class-IL | NCT [40] | ICLR 2023 | 51.59 ± 0.41 | 22.48 ± 19.50 | 26.38 ± 0.57 | 27.40 ± 1.65 | 10.95 ± 1.45 | 49.33 ± 4.47 |
| Class-IL | ProNC [36] | ICLR 2026 | 65.58 ± 0.15 | 32.75 ± 4.71 | 42.99 ± 0.85 | 36.07 ± 0.51 | 27.44 ± 1.00 | 42.81 ± 0.56 |
| Class-IL | Ours | This work | 70.36 ± 0.62 | 22.85 ± 1.45 | 49.74 ± 0.71 | 16.27 ± 0.45 | 31.42 ± 0.80 | 23.84 ± 1.15 |
| Task-IL | ER [29] | ICLR 2019 | 91.19 ± 0.94 | 6.07 ± 1.09 | 60.19 ± 1.01 | 27.38 ± 1.46 | 38.17 ± 2.00 | 40.47 ± 1.54 |
| Task-IL | iCaRL [28] | CVPR 2017 | 88.99 ± 2.13 | 25.34 ± 1.64 | 51.43 ± 1.47 | 36.20 ± 1.85 | 28.19 ± 1.47 | 42.47 ± 2.47 |
| Task-IL | DER [3] | NeurIPS 2020 | 91.40 ± 0.92 | 6.08 ± 0.70 | 63.09 ± 1.09 | 25.98 ± 1.55 | 40.22 ± 0.67 | 40.43 ± 1.05 |
| Task-IL | DER++ [3] | NeurIPS 2020 | 91.92 ± 0.60 | 5.16 ± 0.21 | 66.80 ± 0.41 | 23.91 ± 0.55 | 43.06 ± 1.16 | 39.02 ± 0.43 |
| Task-IL | Co2L [4] | ICCV 2021 | 84.69 ± 1.52 | 2.91 ± 5.25 | 49.19 ± 0.91 | 36.81 ± 0.63 | 37.07 ± 1.62 | 40.75 ± 3.55 |
| Task-IL | CILA [37] | ICML 2024 | 91.36 ± 0.08 | 5.49 ± 0.74 | 53.93 ± 1.02 | 33.07 ± 0.96 | 37.32 ± 1.87 | 41.40 ± 0.51 |
| Task-IL | MNC3L [6] | WACV 2025 | 85.94 ± 0.22 | 4.90 ± 0.15 | 43.91 ± 0.76 | 39.79 ± 0.98 | 32.78 ± 3.54 | 45.10 ± 1.57 |
| Task-IL | STAR [8] | ICLR 2025 | 93.54 ± 1.48 | 6.23 ± 2.06 | 58.53 ± 13.06 | 27.84 ± 8.01 | 41.70 ± 0.95 | 34.78 ± 2.08 |
| Task-IL | CSReL [33] | ICLR 2025 | 69.22 ± 3.03 | 17.16 ± 2.91 | 66.99 ± 0.35 | 23.20 ± 1.74 | 45.04 ± 5.86 | 34.12 ± 2.18 |
| Task-IL | NCT [40] | ICLR 2023 | 80.63 ± 0.46 | 1.41 ± 1.09 | 75.75 ± 0.17 | 4.79 ± 0.07 | 52.71 ± 4.12 | 15.88 ± 0.95 |
| Task-IL | ProNC [36] | ICLR 2026 | 96.86 ± 0.10 | 0.65 ± 0.08 | 85.63 ± 0.73 | 4.40 ± 0.82 | 69.09 ± 0.65 | 9.42 ± 0.58 |
| Task-IL | Ours | This work | 96.67 ± 0.09 | 0.73 ± 0.19 | 87.28 ± 0.25 | 2.36 ± 0.14 | 69.59 ± 0.99 | 8.61 ± 1.09 |
Table 2. Literature comparison for buffer size 200. FAA denotes final average accuracy, and FF denotes final forgetting.
| Setting | Method | Venue | Seq-CIFAR-10 FAA ↑ | Seq-CIFAR-10 FF ↓ | Seq-CIFAR-100 FAA ↑ | Seq-CIFAR-100 FF ↓ | Seq-TinyImageNet FAA ↑ | Seq-TinyImageNet FF ↓ |
| Class-IL | ER [29] | ICLR 2019 | 57.74 ± 2.48 | 43.22 ± 2.10 | 32.35 ± 0.61 | 73.08 ± 0.78 | 9.99 ± 0.29 | 75.21 ± 0.54 |
| Class-IL | iCaRL [28] | CVPR 2017 | 47.55 ± 3.95 | 28.20 ± 2.41 | 33.25 ± 1.25 | 40.99 ± 1.02 | 9.38 ± 1.53 | 37.30 ± 1.42 |
| Class-IL | DER [3] | NeurIPS 2020 | 70.51 ± 1.67 | 24.02 ± 1.63 | 41.36 ± 1.76 | 49.07 ± 2.54 | 17.75 ± 1.14 | 59.95 ± 3.31 |
| Class-IL | DER++ [3] | NeurIPS 2020 | 72.70 ± 1.36 | 22.38 ± 4.41 | 38.20 ± 1.00 | 49.18 ± 2.19 | 19.38 ± 1.41 | 58.75 ± 1.93 |
| Class-IL | Co2L [4] | ICCV 2021 | 61.78 ± 4.22 | 17.79 ± 3.36 | 26.64 ± 1.42 | 48.60 ± 1.34 | 18.71 ± 0.84 | 49.64 ± 1.14 |
| Class-IL | CILA [37] | ICML 2024 | 67.82 ± 0.33 | 18.22 ± 1.32 | 31.27 ± 0.17 | 45.67 ± 1.02 | 19.08 ± 0.49 | 64.14 ± 0.69 |
| Class-IL | MNC3L [6] | WACV 2025 | 52.20 ± 1.56 | 27.88 ± 2.75 | 22.29 ± 0.18 | 46.09 ± 0.58 | 11.52 ± 0.01 | 44.96 ± 0.09 |
| Class-IL | STAR [8] | ICLR 2025 | 69.15 ± 3.53 | 18.59 ± 2.91 | 28.45 ± 1.70 | 53.11 ± 1.57 | 15.19 ± 2.61 | 63.42 ± 3.19 |
| Class-IL | CSReL [33] | ICLR 2025 | 40.82 ± 4.09 | 29.18 ± 6.91 | 39.22 ± 1.70 | 47.03 ± 2.57 | 22.13 ± 0.35 | 50.01 ± 1.87 |
| Class-IL | NCT [40] | ICLR 2023 | 60.93 ± 0.94 | 13.82 ± 3.64 | 33.84 ± 0.38 | 24.37 ± 5.42 | 18.24 ± 0.62 | 50.90 ± 1.16 |
| Class-IL | ProNC [36] | ICLR 2026 | 73.95 ± 0.68 | 26.75 ± 1.29 | 48.94 ± 0.44 | 24.32 ± 0.24 | 29.06 ± 0.32 | 38.58 ± 0.39 |
| Class-IL | Ours | This work | 73.68 ± 0.88 | 17.73 ± 0.88 | 53.27 ± 0.41 | 10.18 ± 0.38 | 32.24 ± 0.05 | 22.47 ± 0.13 |
| Task-IL | ER [29] | ICLR 2019 | 93.61 ± 0.27 | 3.50 ± 0.53 | 73.98 ± 1.52 | 16.23 ± 1.06 | 48.64 ± 0.46 | 30.73 ± 0.62 |
| Task-IL | iCaRL [28] | CVPR 2017 | 88.22 ± 2.62 | 22.61 ± 3.97 | 58.16 ± 1.76 | 27.90 ± 1.37 | 51.55 ± 3.27 | 39.44 ± 0.84 |
| Task-IL | DER [3] | NeurIPS 2020 | 93.40 ± 0.21 | 3.72 ± 0.55 | 71.73 ± 0.74 | 25.98 ± 1.55 | 51.78 ± 0.88 | 28.21 ± 0.97 |
| Task-IL | DER++ [3] | NeurIPS 2020 | 93.88 ± 0.50 | 4.66 ± 1.15 | 74.77 ± 0.31 | 15.75 ± 0.48 | 51.91 ± 0.68 | 25.47 ± 1.03 |
| Task-IL | Co2L [4] | ICCV 2021 | 89.51 ± 2.45 | 2.65 ± 1.00 | 62.32 ± 0.19 | 24.37 ± 0.27 | 50.74 ± 1.24 | 13.80 ± 2.10 |
| Task-IL | CILA [37] | ICML 2024 | 93.29 ± 0.24 | 0.65 ± 0.13 | 68.29 ± 0.46 | 16.98 ± 0.56 | 49.19 ± 0.91 | 27.83 ± 1.32 |
| Task-IL | MNC3L [6] | WACV 2025 | 85.94 ± 0.22 | 3.15 ± 0.42 | 56.43 ± 0.29 | 24.29 ± 0.50 | 36.32 ± 0.05 | 39.00 ± 0.17 |
| Task-IL | STAR [8] | ICLR 2025 | 95.36 ± 0.37 | 2.38 ± 0.34 | 74.06 ± 1.63 | 15.22 ± 0.96 | 55.06 ± 2.07 | 15.19 ± 2.61 |
| Task-IL | CSReL [33] | ICLR 2025 | 73.30 ± 5.92 | 16.89 ± 3.19 | 75.09 ± 0.78 | 15.55 ± 0.76 | 51.94 ± 0.22 | 28.20 ± 0.36 |
| Task-IL | NCT [40] | ICLR 2023 | 81.27 ± 0.24 | 3.84 ± 0.35 | 76.06 ± 0.52 | 3.97 ± 0.75 | 62.30 ± 0.41 | 6.84 ± 0.87 |
| Task-IL | ProNC [36] | ICLR 2026 | 96.95 ± 0.14 | 0.24 ± 0.14 | 86.38 ± 0.43 | 4.35 ± 0.17 | 69.77 ± 0.89 | 9.52 ± 0.96 |
| Task-IL | Ours | This work | 96.81 ± 0.04 | 0.54 ± 0.08 | 87.74 ± 0.07 | 2.02 ± 0.16 | 70.16 ± 0.20 | 7.90 ± 0.16 |
Table 3. Literature comparison for buffer size 500. FAA denotes final average accuracy, and FF denotes final forgetting.
On Seq-CIFAR-100, Class-IL FAA improves by +6.75 points with buffer 200 and +4.33 with buffer 500; Class-IL FF is more than halved in both cases. Seq-TinyImageNet improves under both budgets, and Seq-CIFAR-10/200 gains +4.78 points. The only published-FAA exception is Seq-CIFAR-10/500, where the reported reference is 0.27 points higher. This is the benign regime predicted by the mechanism: only five task blocks compete, Task-IL is near 97%, and the Task-IL–Route gap is smallest. On matched checkpoints, RCR still improves this setting by +1.52 Class-IL FAA, +1.62 Route, and −6.45 Class-IL FF over the reproduced ETF head (Table 1).
Ablation
We ablate the correction on Seq-CIFAR-100 with buffer 200, the hardest CIFAR setting, while keeping the final checkpoint features fixed.
Table 4 isolates the mechanism. Transported radial evidence gives the largest gain over ETF-only scoring. A shared radius removes class-dependent reliability and degrades Route. Removing NC retraction causes a smaller but consistent drop. We therefore interpret the retraction as an NC-consistent stabilization term rather than as the primary source of the gain: the main improvement comes from transport, radial reliability, and the competitive radius. Radial evidence without transport is inert in the composed head: its center is the ETF vertex, so after score normalization it matches the ETF decision.
| Head | ETF | Transport | Radial | Comp. | NC shrink | Class-IL ↑ | Route ↑ |
| ETF only | ✓ | ||||||
| Radial only | ✓ | ✓ | |||||
| No transport | ✓ | ✓ | ✓ | ||||
| Equal radius | ✓ | ✓ | ✓ | ||||
| No NC shrink | ✓ | ✓ | ✓ | ✓ | |||
| Ours (RCR) | ✓ | ✓ | ✓ | ✓ | ✓ |
Table 4. Each component targets routing. Seq-CIFAR-100/200 ablation on the same checkpoint features.
Note: Columns correspond to ETF evidence (Equation (13)), transport (Equations (14) and (15)), radial evidence (Equation (19)), competitive radii (Equation (17)), and NC shrinkage (Equation (18)).
We also test the simplest alternative explanation: a replay-fitted task bias has direct access to replay labels, yet worsens held-out Route (Table 5). This matches Figure 4: replay samples are too clean to expose test-time routing drift.
| Head | Class-IL ↑ | Task-IL ↑ | Route ↑ | Wrong-task ↓ |
| ProNC ETF | 41.71±0.61 | 85.53±0.16 | 42.81±0.57 | 57.19±0.57 |
| Replay task-bias | 29.91±0.64 | 85.53±0.16 | 30.66±0.59 | 69.34±0.59 |
| Ours (RCR) | 49.74±0.58 | 87.28±0.20 | 51.43±0.47 | 48.57±0.47 |
Table 5. Replay-fitted calibration does not explain RCR. A task-bias calibration fit on replay labels is evaluated on held-out Seq-CIFAR-100/200 test features at the final task. It worsens routing, confirming that the replay buffer is not a reliable calibration set.
Full classifier-variant benchmarks and sensitivity studies are in the supplement. We additionally include a small NCT diagnostic to check that the route-correction signal is not tied to one progressive-ETF implementation. On Seq-CIFAR-100 with buffer 200, an offline RCR-style ramp applied to a reproduced NCT-style progressive ETF model improves Class-IL FAA by +2.40, reduces Class-IL FF by 1.21, and improves route accuracy by +2.32 (Table S4). This is a scope diagnostic, not a full benchmark of NCT.
| Head | Class FAA | Class FF | Task FAA | Task FF | Route |
| NCT-style ETF | 31.62 | 27.05 | 79.93 | 4.99 | 33.58 |
| NCT-style ETF + RCR-style ramp | 34.02 | 25.84 | 81.33 | 3.31 | 35.89 |
| +2.40 | +1.41 | +2.32 |
Table S4. Diagnostic check on a second progressive-ETF learner. We reproduce an NCT-style progressive ETF run on Seq-CIFAR-100 with buffer 200 and evaluate an offline RCR-style ramp correction on the same saved features. The ramp keeps ETF evidence dominant at early tasks and increases the radial correction later. The purpose is diagnostic rather than a full re-benchmarking of NCT: it checks that the route-correction signal is not unique to the ProNC implementation.
Mechanistic Analysis
The mechanism has two parts. The ETF score supplies a globally comparable NC frame, while the transported radial score asks whether is locally plausible under the current class measure. Score normalization makes the two evidences commensurate before composition. In the exact NC limit, the radial score is a positive affine transform of the ETF score, so RCR changes no decisions.
| Dataset | Buf. | NC1 disp. | CV() | Task–Route gap | Route | Class | |
| Seq-CIFAR-10 | 200 | 1.50±0.01 | 0.37±0.02 | 0.13±0.00 | 30.16±1.24 | +4.23±1.41 | +4.22±1.40 |
| Seq-CIFAR-10 | 500 | 1.44±0.01 | 0.38±0.02 | 0.14±0.01 | 24.11±1.36 | +1.62±1.14 | +1.52±1.05 |
| Seq-CIFAR-100 | 200 | 1.61±0.00 | 0.41±0.00 | 0.11±0.00 | 42.71±0.41 | +8.62±0.11 | +8.03±0.19 |
| Seq-CIFAR-100 | 500 | 1.52±0.00 | 0.43±0.00 | 0.10±0.00 | 35.41±0.47 | +4.90±0.32 | +4.62±0.40 |
| Seq-TinyImageNet | 200 | 1.63±0.00 | 0.43±0.00 | 0.08±0.00 | 39.72±0.65 | +6.18±0.72 | +5.65±0.69 |
| Seq-TinyImageNet | 500 | 1.60±0.00 | 0.43±0.00 | 0.07±0.00 | 38.33±0.73 | +6.09±0.36 | +5.63±0.35 |
Table 6. RCR is most active when routing drift is largest. In the exact NC limit, transported class representatives coincide with ETF vertices and RCR reduces to nearest-ETF inference. Across validated runs, the head is least active when the ProNC task-local classifier already routes well, and its gains grow with the held-out Task-IL–Route gap (row-wise with Route and with Class). Values are final-task mean ± std over three seeds.
Note: is the mean cosine distance between transported class representatives and ETF vertices. Task–Route gap is ProNC Task-IL FAA minus ProNC route accuracy.
Table 6 makes the link quantitative across datasets: the smallest gains occur when the ProNC task-local classifier already routes well, and gains grow with the Task-IL–Route gap. Detailed stage-level NC diagnostics, class-wise correlations, and true-route headroom are reported in the supplementary material (Table S11 and Table S12).
Table S11. Neural-collapse deviations and routing drift on Seq-CIFAR-100, buffer 200. Panel (a) tracks stage-level NC diagnostics and held-out performance. Panel (b) reports final-stage class-wise Pearson correlations: prototype–ETF deviation and competitive radius identify older classes with larger route error and larger RCR gains, while NC1 dispersion mostly tracks age. Values are mean std over saved runs.
| Stage | ↑ | NC1 disp. ↓ | CV() | Route ↑ | Class-IL ↑ |
| T3 | 0.390.00 | 0.270.00 | 0.080.01 | 75.341.01 74.790.83 | 70.980.89 70.580.63 |
| T10 | 0.190.00 | 0.410.00 | 0.110.00 | 42.810.57 51.430.47 | 41.710.61 49.740.58 |
| (b) Final-stage class-wise correlations | |||||
| Class signal | Route error | Class | Age | ||
| 0.680.02 | 0.590.03 | 0.900.02 | |||
| 0.610.00 | 0.550.05 | 0.940.01 | |||
| NC1 disp. | 0.240.03 | 0.130.04 | 0.810.01 |
caption: Note: NC1 dispersion is the held-out mean within each class. Route error and Class are computed per class before correlation.
Table S12. Route-oracle bounds.
| Dataset/Buffer | Head | Class-IL ↑ | Task-IL ↑ | Route ↑ |
| Seq-CIFAR-10 / 200 | ProNC ETF | 66.141.38 | 96.740.10 | 66.581.34 |
| RCR | 70.360.50 | 96.670.08 | 70.820.46 | |
| ProNC + true route | 96.740.10 | 96.740.10 | 100.000.00 | |
| RCR + true route | 96.670.08 | 96.670.08 | 100.000.00 | |
| Seq-CIFAR-10 / 500 | ProNC ETF | 72.161.30 | 96.820.12 | 72.711.35 |
| RCR | 73.680.72 | 96.810.03 | 74.330.83 | |
| ProNC + true route | 96.820.12 | 96.820.12 | 100.000.00 | |
| RCR + true route | 96.810.03 | 96.810.03 | 100.000.00 | |
| Seq-CIFAR-100 / 200 | ProNC ETF | 41.710.61 | 85.530.16 | 42.810.57 |
| RCR | 49.740.58 | 87.280.20 | 51.430.47 | |
| ProNC + true route | 85.530.16 | 85.530.16 | 100.000.00 | |
| RCR + true route | 87.280.20 | 87.280.20 | 100.000.00 | |
| Seq-CIFAR-100 / 500 | ProNC ETF | 48.650.55 | 85.540.19 | 50.140.42 |
| RCR | 53.270.34 | 87.740.05 | 55.040.21 | |
| ProNC + true route | 85.540.19 | 85.540.19 | 100.000.00 | |
| RCR + true route | 87.740.05 | 87.740.05 | 100.000.00 | |
| Seq-TinyImageNet / 200 | ProNC ETF | 25.770.41 | 68.300.62 | 28.580.31 |
| RCR | 31.420.65 | 69.590.81 | 34.760.56 | |
| ProNC + true route | 68.300.62 | 68.300.62 | 100.000.00 | |
| RCR + true route | 69.590.81 | 69.590.81 | 100.000.00 | |
| Seq-TinyImageNet / 500 | ProNC ETF | 26.610.32 | 70.310.39 | 29.370.42 |
| RCR | 32.240.04 | 70.160.16 | 35.460.07 | |
| ProNC + true route | 67.710.39 | 67.710.39 | 100.000.00 | |
| RCR + true route | 70.160.16 | 70.160.16 | 100.000.00 |
caption: Note: The true-route rows restrict the argmax to the ground-truth task block and are upper bounds, not deployable methods.
Seq-CIFAR-10 is closest to the benign regime: fewer task blocks compete, Task-IL is near 97%, and route accuracy is already high. Accordingly, RCR gives smaller gains there, especially with buffer 500. Gains grow when the Task-IL–Route gap is larger, supporting the intended interpretation: RCR is an NC-deviation correction, not a new within-task classifier. The supplement further shows remaining true-route headroom and class-wise links between center drift, radius growth, class age, and route error.
Discussion
RCR targets the route-limited regime: the representation remains useful inside the correct task block, but the global score no longer compares task blocks reliably. It is not a universal repair for arbitrary forgetting; if Task-IL is low, the within-block class structure has already been lost. Its classifier-level scope is useful because the backbone and ETF frame stay fixed: the gains show that trained progressive NC models contain class-measure information that the raw ETF score underuses. RCR changes class evidence, not the evaluation protocol, and does not insert a learned task oracle.
This also clarifies the role of RCR in the paper. The contribution is not a large auxiliary learner on top of ProNC, but a controlled test of whether the NC geometry already learned by the model is being scored correctly. If the raw ETF classifier were already using all useful geometric information, transported class centers and radii would not improve held-out routing. The observed gains therefore indicate a specific mismatch: the global ETF frame remains valuable, but the local class measure has drifted away from the ideal point prototype.
Replay is used only to estimate class-local concentration around transported representatives. This differs from fitting a replay task bias, which has label access but transfers poorly to held-out routing. Inference adds one prototype-distance score matrix, per-sample standardization, and no extra backbone forward; for TinyImageNet, the class-level state is about 0.39 MB in FP32. The route-oracle analysis in the supplement should therefore be read constructively: it does not assume a deployable task oracle, but measures how much of the remaining error is still attributable to block selection rather than within-block discrimination. The remaining gap motivates training-time objectives that make representations route-stable rather than only corrected at inference.
Conclusion
We isolate routing drift as a measurable failure mode of progressive neural-collapse continual learning: Task-IL can remain high while Class-IL fails because the global argmax routes samples to the wrong task block. RCR addresses this regime by preserving the trained ETF frame and adding transported radial reliability, improving Class-IL and forgetting while preserving Task-IL. Across the validated settings, the same fixed correction sets the best Class-IL FAA in five of six cases and lowers forgetting in all six. This positions neural collapse as both a training target and a diagnostic geometry for score-comparability failures.
Supplementary Material
Supplement Overview
This supplement is organized as a companion to the main paper. We first record the implementation details needed to reproduce the trained artifacts, including the reference ProNC objective and hyperparameters. We then collect notation and neural-collapse background, and define the classifier variants evaluated from the same frozen features and saved task artifacts. The remaining sections separate three questions: which variant is strongest, which components matter, and whether the observed gains match the routing-drift mechanism. Throughout, RCR denotes the deployed Reliability-Calibrated Routing score; other heads are controlled variants or ablations.
Implementation Details
This section records the training and evaluation details behind the frozen artifacts used in the main paper. Our contribution is evaluated as a classifier correction on top of a trained progressive ETF model; the backbone checkpoints are produced with the reference ProNC objective [36], while RCR is applied after training from saved task artifacts.
Reference training objective. The first task is trained with ordinary supervised cross-entropy. At later tasks, the ETF frame is expanded before learning the new task, and each update uses a minibatch containing current-task samples and replay samples. Let be the normalized current feature, the normalized feature from the frozen previous model, and the progressive ETF vertex assigned to label at stage . The training loss is
The alignment term is the squared cosine distance-to-one objective
which pulls the current normalized feature toward the ETF vertex of its class. The feature-distillation term is
evaluated on the same mixed minibatch. Thus, controls the strength of ETF alignment and controls the strength of feature preservation from the previous model. The cross-entropy term keeps its unit coefficient in all reproduced runs.
Training hyperparameters. All reproduced artifacts use a ResNet-18 backbone, batch size 32, SGD with zero momentum and zero weight decay, and a multistep learning-rate schedule with decay factor 0.1. The class order is fixed by the benchmark split, and all reported paired results average seeds 0, 1, 2. The dataset-specific values used to produce the saved checkpoints are listed in Table S1.
| Dataset | Buffer | Epochs | LR | Milestones | ||
| Seq-CIFAR-10 | 200 | 50 | 0.01 | 35, 45 | 13 | 90 |
| Seq-CIFAR-10 | 500 | 50 | 0.01 | 35, 45 | 12 | 80 |
| Seq-CIFAR-100 | 200 | 50 | 0.03 | 35, 45 | 18 | 170 |
| Seq-CIFAR-100 | 500 | 50 | 0.03 | 35, 45 | 18 | 170 |
| Seq-TinyImageNet | 200 | 100 | 0.03 | 45, 75, 95 | 20 | 165 |
| Seq-TinyImageNet | 500 | 100 | 0.03 | 45, 75, 95 | 20 | 165 |
Table S1. Training hyperparameters for reproduced artifacts. The same checkpoints are used to evaluate both the reference ETF head and RCR in paired comparisons.
Saved artifacts and post-hoc evaluation. At each task boundary we save the checkpoint, the expanded ETF frame, replay features needed for class representatives, and the previous-to-current feature transport fit. The reference ETF head predicts with
RCR uses the same frozen feature and the same ETF frame. It adds only the transported class representative, competitive radius, NC retraction, and fixed ETF/radial composition defined in the main paper. No scalar is chosen per dataset or from a validation split: all datasets use , , shrink clipping to , radius clipping to , and .
Transport fitting. Class representatives are moved across feature frames by fitting a ridge linear map between previous-model and current-model normalized features from the mixed training stream at a task boundary. The reported artifacts use ridge coefficient , pair normalization before fitting, and normalized features at evaluation. The transported representative is therefore a class-level measure in the current feature frame, while remains the global ETF vertex used by the neural-collapse classifier.
Notation and Neural-Collapse Background
Table S2 summarizes the notation used throughout the paper and this supplement. We keep protocol symbols, routing diagnostics, and RCR components separate to avoid overloading task indices, class sets, and prototype quantities.
We recall the standard neural-collapse terminology used in the main paper. Consider a balanced -class classification problem. Let be the last-layer feature of sample from class , the class mean, the global mean, and the classifier weight for class . Neural collapse describes the terminal training regime in which the following four properties emerge [27].
NC1: within-class variability collapse. Features from the same class concentrate around their class mean:
NC2: simplex ETF structure of class means. The centered class means
approach a simplex equiangular tight frame, so all classes have equal norm and equal pairwise angle:
NC3: self-duality. Classifier weights align with centered class means:
NC4: nearest-class-center inference. Under NC1–NC3, the classifier decision is equivalent to nearest-class-center or nearest-prototype inference in the collapsed feature space. This is the connection exploited by ETF classifiers: once the class directions form a simplex ETF, prediction can be performed by comparing a normalized feature to class vertices through cosine similarity.
Offline Classifier Variants
All variants below are evaluated on the same frozen feature tensor for a given run. We use for the normalized feature, for the ProNC ETF vertex of class , and for the transported class prototype in the current feature frame:
We write for per-sample standardization across the currently seen classes. Unless stated otherwise, no variant uses test-time task identity or a learned task oracle.
Evaluation artifacts. All variants are evaluated from saved task and feature artifacts. The same saved features are used for ProNC ETF, LDC-NCM, radial scoring, and reliability-calibrated scoring. This isolates the classifier effect from training variance.
ETF reference. This is the nearest-ETF score used by the base progressive neural-collapse classifier:
This is the baseline head in the main tables.
CE head (audit). The trained linear classifier head uses its saved weights and bias. We report it only as an audit row, because the controlled setting evaluates neural-collapse heads with ETF inference rather than this linear classifier.
Transported NCM. This variant uses transported nearest-centroid inference:
It tests whether moving class representatives into the current feature frame is useful without adding radial reliability.
Radial-only. This variant scores local plausibility by normalizing the cosine distance to with a class radius:
It tests whether local class reliability alone can repair cross-task routing.
CompMix. CompMix composes the ETF score with the competitive-radius radial score around the transported prototype:
The deterministic stage schedule keeps the radial term weak at early tasks and approaches the fixed cap as more task blocks are observed.
CompMix-Fixed. CompMix-Fixed uses the same competitive-radius radial score around , but removes the stage schedule and uses the fixed cap at every task:
It isolates whether the gain comes from the evidence composition itself or from a deterministic task-stage schedule.
InvGradMix (directional sensitivity). InvGradMix replaces the Euclidean radial distance with a global diagonal metric estimated from replay samples using inverse empirical CE-gradient magnitude:
It then uses the same stage schedule as CompMix. This variant tests whether different feature directions should have different reliability.
InvGradMix-Fixed (directional sensitivity). InvGradMix-Fixed uses the same direction-aware radial score as InvGradMix, but uses the fixed cap instead of the stage schedule. It is a stronger direction-aware variant, but it introduces metric-estimation choices that are not needed by RCR.
NC-ShrinkMix. NC-ShrinkMix adds the neural-collapse retraction to CompMix. It first retracts the transported prototype toward the ETF vertex,
with . It then recomputes the competitive class radius and combines ETF evidence with radial evidence using the scheduled mixture .
RCR. RCR is the fixed-cap version of NC-ShrinkMix. It uses the retracted prototype , recomputes the competitive radius around that prototype, and combines ETF evidence with radial evidence using :
This is the method used in all main-paper results.
Saved mix (audit). Some artifacts contain the mixture that was active when the run was produced. Because this score can depend on historical choices of auxiliary head, radius mode, and mixture weight, we report it only as an audit row. It is not a method candidate and is included only to make older saved artifacts auditable.
No class-index prior. The proposed score does not use the numerical class index, task age, or a hand-coded penalty such as or in the logits. The only mixture quantity in RCR is the global scalar , shared by every class and sample. Class-specific terms are data-derived representatives and uncertainties, , exactly as in prototype-based classifiers. The task index is used for evaluation metrics such as Task-IL and route accuracy, but not as a test-time oracle or as a per-class routing bias.
Classifier Variant Benchmark
The main paper reports the paired ETF-versus-RCR comparison. Here we show the broader classifier family on the same frozen artifacts. The purpose is not to introduce multiple methods, but to verify which ingredient explains the gains. RCR is the deployed variant because it combines ETF evidence, transported class reliability, and NC retraction without the additional metric-estimation choices used by the direction-aware variants. Rows explicitly marked as audits are diagnostic controls rather than candidate methods; they are included to show which apparent shortcuts fail under the same frozen-feature evaluation.
Analysis. Table S3 shows three consistent patterns. First, the audit CE head is not a competitive replacement for ETF inference: it often preserves Task-IL but collapses Class-IL, confirming that the comparison should be made between neural-collapse heads on the same feature artifacts. Second, transported NCM and radial-only scoring are useful but incomplete: they improve hard routing regimes such as Seq-CIFAR-100 and Seq-TinyImageNet, yet either underperform on benign Seq-CIFAR-10 or lose the global ETF ordering. Third, the large gains appear when the ETF score and local radial reliability are composed. CompMix and NC-ShrinkMix therefore separate the two useful signals: global ETF comparability and local class plausibility. The fixed variants remove the deterministic stage schedule; RCR is selected because it gives the most stable forgetting reduction while keeping the isotropic NC story intact. The direction-aware InvGradMix variants are deliberately left as ablations: they show that additional feature-direction reliability can sometimes improve FAA or Route, but at the cost of an extra learned metric and a less parsimonious method.
| Dataset | Buffer | Head | Class FAA | Class FF | Task FAA | Task FF | Route | |
| CIFAR-10 | 200 | ETF reference | 3 | 66.14 ± 1.69 | 31.47 ± 2.46 | 96.74 ± 0.12 | 0.62 ± 0.20 | 66.58 ± 1.64 |
| CE head (audit) | 3 | 54.63 ± 0.15 | 52.99 ± 0.03 | 96.62 ± 0.24 | 0.75 ± 0.24 | 55.04 ± 0.09 | ||
| Transported NCM | 3 | 63.46 ± 0.75 | 20.75 ± 1.21 | 96.11 ± 0.53 | 1.36 ± 0.66 | 63.98 ± 0.66 | ||
| Radial-only | 3 | 64.38 ± 1.23 | 15.80 ± 1.84 | 96.11 ± 0.53 | 1.36 ± 0.66 | 65.06 ± 0.98 | ||
| CompMix | 3 | 69.84 ± 0.76 | 24.39 ± 1.40 | 96.63 ± 0.10 | 0.75 ± 0.22 | 70.30 ± 0.80 | ||
| CompMix-Fixed | 3 | 69.84 ± 0.76 | 24.15 ± 1.46 | 96.63 ± 0.10 | 0.77 ± 0.22 | 70.30 ± 0.80 | ||
| InvGradMix | 3 | 70.65 ± 0.71 | 22.59 ± 1.17 | 96.72 ± 0.09 | 0.67 ± 0.14 | 71.11 ± 0.73 | ||
| InvGradMix-Fixed | 3 | 70.65 ± 0.71 | 22.16 ± 1.18 | 96.72 ± 0.09 | 0.69 ± 0.13 | 71.11 ± 0.73 | ||
| NC-ShrinkMix | 3 | 70.36 ± 0.62 | 23.03 ± 1.40 | 96.67 ± 0.09 | 0.72 ± 0.19 | 70.82 ± 0.56 | ||
| RCR | 3 | 70.36 ± 0.62 | 22.85 ± 1.45 | 96.67 ± 0.09 | 0.73 ± 0.19 | 70.82 ± 0.56 | ||
| Saved mix (audit) | 3 | 68.41 ± 0.58 | 26.72 ± 0.93 | 96.59 ± 0.30 | 0.80 ± 0.38 | 68.93 ± 0.51 | ||
| CIFAR-10 | 500 | ETF reference | 3 | 72.16 ± 1.59 | 24.18 ± 2.14 | 96.82 ± 0.14 | 0.47 ± 0.04 | 72.71 ± 1.65 |
| CE head (audit) | 3 | 65.60 ± 0.48 | 38.90 ± 0.48 | 96.86 ± 0.15 | 0.51 ± 0.10 | 66.03 ± 0.48 | ||
| Transported NCM | 3 | 66.28 ± 0.34 | 18.60 ± 0.53 | 96.34 ± 0.04 | 0.97 ± 0.13 | 66.95 ± 0.36 | ||
| Radial-only | 3 | 67.81 ± 1.05 | 12.52 ± 1.58 | 96.34 ± 0.04 | 0.97 ± 0.13 | 68.51 ± 0.96 | ||
| CompMix | 3 | 73.56 ± 0.04 | 18.42 ± 0.85 | 96.76 ± 0.01 | 0.59 ± 0.12 | 74.22 ± 0.19 | ||
| CompMix-Fixed | 3 | 73.56 ± 0.04 | 18.35 ± 0.82 | 96.76 ± 0.01 | 0.59 ± 0.13 | 74.22 ± 0.19 | ||
| InvGradMix | 3 | 74.00 ± 0.19 | 17.25 ± 0.94 | 96.86 ± 0.09 | 0.48 ± 0.01 | 74.65 ± 0.32 | ||
| InvGradMix-Fixed | 3 | 74.00 ± 0.19 | 17.06 ± 1.01 | 96.86 ± 0.09 | 0.49 ± 0.02 | 74.65 ± 0.32 | ||
| NC-ShrinkMix | 3 | 73.68 ± 0.88 | 17.78 ± 0.90 | 96.81 ± 0.04 | 0.54 ± 0.07 | 74.33 ± 1.02 | ||
| RCR | 3 | 73.68 ± 0.88 | 17.73 ± 0.88 | 96.81 ± 0.04 | 0.54 ± 0.08 | 74.33 ± 1.02 | ||
| Saved mix (audit) | 3 | 72.90 ± 0.56 | 19.97 ± 0.76 | 96.79 ± 0.01 | 0.52 ± 0.15 | 73.53 ± 0.49 | ||
| CIFAR-100 | 200 | ETF reference | 3 | 41.71 ± 0.75 | 38.55 ± 1.07 | 85.53 ± 0.20 | 4.51 ± 0.40 | 42.81 ± 0.70 |
| CE head (audit) | 3 | 19.77 ± 0.51 | 77.33 ± 0.88 | 86.32 ± 0.10 | 3.76 ± 0.20 | 20.48 ± 0.51 | ||
| Transported NCM | 3 | 43.59 ± 0.79 | 31.24 ± 1.00 | 85.64 ± 0.09 | 4.03 ± 0.17 | 45.41 ± 0.78 | ||
| Radial-only | 3 | 46.78 ± 0.79 | 21.47 ± 0.80 | 85.64 ± 0.09 | 4.03 ± 0.17 | 48.92 ± 0.65 | ||
| CompMix | 3 | 49.28 ± 0.50 | 21.13 ± 0.46 | 87.04 ± 0.21 | 2.76 ± 0.29 | 50.99 ± 0.38 | ||
| CompMix-Fixed | 3 | 49.20 ± 0.48 | 18.24 ± 0.47 | 87.00 ± 0.18 | 2.62 ± 0.18 | 50.93 ± 0.37 | ||
| InvGradMix | 3 | 49.55 ± 0.60 | 20.76 ± 0.68 | 87.17 ± 0.22 | 2.57 ± 0.17 | 51.26 ± 0.50 | ||
| InvGradMix-Fixed | 3 | 49.46 ± 0.51 | 18.00 ± 0.50 | 87.13 ± 0.19 | 2.47 ± 0.17 | 51.21 ± 0.42 | ||
| NC-ShrinkMix | 3 | 49.72 ± 0.66 | 19.31 ± 0.62 | 87.37 ± 0.28 | 2.45 ± 0.22 | 51.38 ± 0.50 | ||
| RCR | 3 | 49.74 ± 0.71 | 16.27 ± 0.45 | 87.28 ± 0.25 | 2.36 ± 0.14 | 51.43 ± 0.57 | ||
| Saved mix (audit) | 3 | 44.53 ± 5.92 | 33.58 ± 16.01 | 87.15 ± 0.36 | 2.57 ± 0.12 | 46.01 ± 6.23 | ||
| CIFAR-100 | 500 | ETF reference | 3 | 48.65 ± 0.68 | 23.16 ± 1.51 | 85.54 ± 0.23 | 4.29 ± 0.31 | 50.14 ± 0.51 |
| CE head (audit) | 3 | 28.56 ± 0.42 | 66.87 ± 0.55 | 86.27 ± 0.25 | 3.85 ± 0.11 | 29.39 ± 0.46 | ||
| Transported NCM | 3 | 50.95 ± 0.31 | 21.25 ± 0.48 | 86.94 ± 0.19 | 2.63 ± 0.19 | 52.83 ± 0.22 | ||
| Radial-only | 3 | 51.96 ± 0.06 | 13.84 ± 0.18 | 86.94 ± 0.19 | 2.63 ± 0.19 | 53.94 ± 0.07 | ||
| CompMix | 3 | 53.15 ± 0.49 | 12.71 ± 0.26 | 87.55 ± 0.15 | 2.30 ± 0.05 | 54.95 ± 0.34 | ||
| CompMix-Fixed | 3 | 53.07 ± 0.41 | 11.44 ± 0.26 | 87.51 ± 0.15 | 2.19 ± 0.11 | 54.86 ± 0.24 | ||
| InvGradMix | 3 | 53.29 ± 0.40 | 12.76 ± 0.30 | 87.58 ± 0.07 | 2.30 ± 0.09 | 55.11 ± 0.25 | ||
| InvGradMix-Fixed | 3 | 53.24 ± 0.38 | 11.42 ± 0.24 | 87.59 ± 0.07 | 2.17 ± 0.13 | 55.07 ± 0.20 | ||
| NC-ShrinkMix | 3 | 53.29 ± 0.49 | 11.32 ± 0.26 | 87.76 ± 0.07 | 2.14 ± 0.12 | 55.04 ± 0.32 | ||
| RCR | 3 | 53.27 ± 0.41 | 10.18 ± 0.38 | 87.74 ± 0.07 | 2.02 ± 0.16 | 55.04 ± 0.26 | ||
| Saved mix (audit) | 3 | 50.23 ± 1.17 | 24.12 ± 14.59 | 87.71 ± 0.08 | 2.00 ± 0.25 | 51.88 ± 4.37 | ||
| TinyImageNet | 200 | ETF reference | 3 | 25.77 ± 0.50 | 43.45 ± 0.13 | 68.30 ± 0.76 | 9.58 ± 0.42 | 28.58 ± 0.38 |
| CE head (audit) | 3 | 9.91 ± 0.09 | 75.27 ± 0.33 | 71.98 ± 0.35 | 6.52 ± 0.09 | 12.20 ± 0.10 | ||
| Transported NCM | 3 | 28.71 ± 0.40 | 33.37 ± 0.45 | 68.69 ± 0.38 | 9.65 ± 0.40 | 32.17 ± 0.45 | ||
| Radial-only | 3 | 30.46 ± 0.20 | 26.66 ± 0.72 | 68.69 ± 0.38 | 9.65 ± 0.40 | 33.99 ± 0.16 | ||
| CompMix | 3 | 30.68 ± 0.78 | 30.89 ± 0.88 | 69.79 ± 0.99 | 8.35 ± 1.02 | 33.96 ± 0.68 | ||
| CompMix-Fixed | 3 | 30.18 ± 0.92 | 26.84 ± 1.29 | 68.54 ± 0.98 | 9.62 ± 1.03 | 33.74 ± 0.90 | ||
| InvGradMix | 3 | 30.36 ± 0.89 | 31.20 ± 0.95 | 69.75 ± 0.90 | 8.42 ± 0.93 | 33.59 ± 0.87 | ||
| InvGradMix-Fixed | 3 | 29.92 ± 0.91 | 27.41 ± 1.12 | 68.53 ± 0.95 | 9.76 ± 1.04 | 33.38 ± 0.86 | ||
| NC-ShrinkMix | 3 | 31.55 ± 0.72 | 28.41 ± 0.83 | 70.55 ± 0.94 | 7.57 ± 0.88 | 34.70 ± 0.55 | ||
| RCR | 3 | 31.42 ± 0.80 | 23.84 ± 1.15 | 69.59 ± 0.99 | 6.81 ± 1.09 | 34.76 ± 0.68 | ||
| Saved mix (audit) | 3 | 31.24 ± 0.44 | 27.95 ± 0.54 | 70.23 ± 0.51 | 8.07 ± 0.48 | 34.53 ± 0.44 | ||
| TinyImageNet | 500 | ETF reference | 3 | 26.61 ± 0.39 | 41.16 ± 0.23 | 67.71 ± 0.47 | 9.78 ± 0.52 | 29.37 ± 0.52 |
| CE head (audit) | 3 | 10.81 ± 0.33 | 73.80 ± 0.38 | 71.59 ± 0.44 | 6.69 ± 0.60 | 13.22 ± 0.28 | ||
| Transported NCM | 3 | 29.59 ± 0.46 | 32.10 ± 0.40 | 69.37 ± 0.43 | 8.71 ± 0.40 | 32.92 ± 0.29 | ||
| Radial-only | 3 | 31.23 ± 0.08 | 25.38 ± 0.24 | 69.37 ± 0.43 | 8.71 ± 0.40 | 34.67 ± 0.02 | ||
| CompMix | 3 | 31.92 ± 0.01 | 28.72 ± 0.35 | 70.32 ± 0.32 | 7.71 ± 0.37 | 35.16 ± 0.15 | ||
| CompMix-Fixed | 3 | 31.34 ± 0.15 | 25.20 ± 0.25 | 69.21 ± 0.20 | 8.91 ± 0.20 | 34.71 ± 0.19 | ||
| InvGradMix | 3 | 31.68 ± 0.12 | 29.10 ± 0.53 | 70.47 ± 0.41 | 7.68 ± 0.50 | 34.91 ± 0.05 | ||
| InvGradMix-Fixed | 3 | 31.09 ± 0.21 | 25.78 ± 0.18 | 69.32 ± 0.18 | 8.87 ± 0.17 | 34.44 ± 0.11 | ||
| NC-ShrinkMix | 3 | 32.47 ± 0.17 | 26.67 ± 0.23 | 70.89 ± 0.44 | 7.10 ± 0.56 | 35.63 ± 0.27 | ||
| RCR | 3 | 32.24 ± 0.05 | 22.47 ± 0.13 | 70.16 ± 0.20 | 7.90 ± 0.16 | 35.46 ± 0.08 | ||
| Saved mix (audit) | 3 | 31.93 ± 0.13 | 26.80 ± 0.36 | 70.53 ± 0.42 | 7.48 ± 0.40 | 35.29 ± 0.12 |
Table S3. Classifier variants evaluated on frozen artifacts. All variants use the same feature artifacts and test split. Route is final route accuracy. Bold marks the best mean and underline marks the second best mean within each dataset/buffer block. Saved mix is included as an audit row because it depends on the score stored at artifact generation time.
Second Progressive-ETF Diagnostic
Reviewer feedback asked whether the route-correction mechanism is tied to one progressive-ETF implementation. To address this without overclaiming, we ran a small diagnostic on an NCT-style progressive ETF model and applied an offline RCR-style ramp correction to the same frozen features. This is not intended as a full NCT reproduction or a replacement for the controlled ProNC comparison in the main paper. It simply checks whether the same route-repair signal appears when the base learner is changed within the NC/ETF family.
Component and Sensitivity Ablations
Score Components
Note: All heads use the same frozen features. No-transport uses the ETF vertex as the radial center, isolating the role of transport-class measures.
Analysis. Table S5 uses the same frozen evaluation artifacts as the main results and separates ETF evidence, transported centroids, radial evidence, competitive radii, and transport. It reveals two failure modes. Transported NCM usually improves over ETF on hard routing regimes, so class representatives do benefit from being moved into the current feature frame. However, NCM alone does not supply a reliability model and can underperform on easier regimes. Conversely, ETF+no-transport collapses back to the ETF reference, showing that radial evidence only matters when it is attached to transported class measures. The full RCR row is strongest because it keeps the global ETF ordering while adding local class plausibility and competitive uncertainty.
| Dataset/Buffer | Head | Class-IL ↑ | Task-IL ↑ | Route ↑ |
| Seq-CIFAR-10 / 200 | ETF only | 66.14 ± 1.38 | 96.74 ± 0.10 | 66.58 ± 1.34 |
| Transported NCM | 63.46 ± 0.61 | 96.11 ± 0.43 | 63.98 ± 0.54 | |
| Radial only | 64.06 ± 0.42 | 96.27 ± 0.15 | 64.66 ± 0.32 | |
| ETF+NCM | 66.31 ± 0.82 | 96.55 ± 0.26 | 66.80 ± 0.70 | |
| ETF+raw radial | 68.50 ± 0.33 | 96.62 ± 0.08 | 69.05 ± 0.29 | |
| ETF+equal radius | 66.31 ± 0.82 | 96.55 ± 0.26 | 66.80 ± 0.70 | |
| ETF+no transport | 66.14 ± 1.38 | 96.76 ± 0.09 | 66.58 ± 1.34 | |
| Ours (RCR) | 70.36 ± 0.50 | 96.67 ± 0.08 | 70.82 ± 0.46 | |
| Seq-CIFAR-10 / 500 | ETF only | 72.16 ± 1.30 | 96.82 ± 0.12 | 72.71 ± 1.35 |
| Transported NCM | 66.28 ± 0.27 | 96.34 ± 0.03 | 66.95 ± 0.29 | |
| Radial only | 68.90 ± 0.81 | 96.24 ± 0.02 | 69.65 ± 0.69 | |
| ETF+NCM | 71.73 ± 0.69 | 96.72 ± 0.04 | 72.32 ± 0.61 | |
| ETF+raw radial | 71.90 ± 1.64 | 96.73 ± 0.13 | 72.61 ± 1.65 | |
| ETF+equal radius | 71.73 ± 0.69 | 96.72 ± 0.04 | 72.32 ± 0.61 | |
| ETF+no transport | 72.16 ± 1.30 | 96.83 ± 0.11 | 72.71 ± 1.35 | |
| Ours (RCR) | 73.68 ± 0.72 | 96.81 ± 0.03 | 74.33 ± 0.83 | |
| Seq-CIFAR-100 / 200 | ETF only | 41.71 ± 0.61 | 85.53 ± 0.16 | 42.81 ± 0.57 |
| Transported NCM | 43.59 ± 0.65 | 85.64 ± 0.07 | 45.41 ± 0.64 | |
| Radial only | 46.21 ± 0.44 | 85.46 ± 0.26 | 48.36 ± 0.49 | |
| ETF+NCM | 44.49 ± 0.29 | 86.64 ± 0.18 | 46.04 ± 0.28 | |
| ETF+raw radial | 43.10 ± 1.07 | 86.57 ± 0.42 | 44.87 ± 1.05 | |
| ETF+equal radius | 44.49 ± 0.29 | 86.64 ± 0.18 | 46.04 ± 0.28 | |
| ETF+no transport | 41.71 ± 0.61 | 85.56 ± 0.16 | 42.81 ± 0.57 | |
| Ours (RCR) | 49.74 ± 0.58 | 87.28 ± 0.20 | 51.43 ± 0.47 | |
| Seq-CIFAR-100 / 500 | ETF only | 48.65 ± 0.55 | 85.54 ± 0.19 | 50.14 ± 0.42 |
| Transported NCM | 50.95 ± 0.25 | 86.94 ± 0.15 | 52.83 ± 0.18 | |
| Radial only | 50.93 ± 0.22 | 86.65 ± 0.10 | 52.88 ± 0.19 | |
| ETF+NCM | 51.47 ± 0.18 | 87.48 ± 0.12 | 53.26 ± 0.10 | |
| ETF+raw radial | 45.98 ± 1.03 | 87.09 ± 0.11 | 47.66 ± 0.96 | |
| ETF+equal radius | 51.47 ± 0.18 | 87.48 ± 0.12 | 53.26 ± 0.10 | |
| ETF+no transport | 48.65 ± 0.55 | 85.61 ± 0.18 | 50.14 ± 0.42 | |
| Ours (RCR) | 53.27 ± 0.34 | 87.74 ± 0.05 | 55.04 ± 0.21 | |
| Seq-TinyImageNet / 200 | ETF only | 25.77 ± 0.41 | 68.30 ± 0.62 | 28.58 ± 0.31 |
| Transported NCM | 28.71 ± 0.33 | 68.69 ± 0.31 | 32.17 ± 0.37 | |
| Radial only | 28.21 ± 0.42 | 66.09 ± 1.00 | 32.14 ± 0.33 | |
| ETF+NCM | 29.29 ± 0.41 | 70.01 ± 0.35 | 32.63 ± 0.37 | |
| ETF+raw radial | 19.96 ± 2.18 | 54.13 ± 2.77 | 24.79 ± 1.02 | |
| ETF+equal radius | 29.29 ± 0.41 | 70.01 ± 0.35 | 32.63 ± 0.37 | |
| ETF+no transport | 25.77 ± 0.41 | 68.31 ± 0.68 | 28.58 ± 0.31 | |
| Ours (RCR) | 31.42 ± 0.65 | 69.59 ± 0.81 | 34.76 ± 0.56 | |
| Seq-TinyImageNet / 500 | ETF only | 26.61 ± 0.32 | 67.71 ± 0.39 | 29.37 ± 0.42 |
| Transported NCM | 29.59 ± 0.37 | 69.37 ± 0.35 | 32.92 ± 0.24 | |
| Radial only | 30.00 ± 0.19 | 67.33 ± 0.13 | 33.72 ± 0.24 | |
| ETF+NCM | 29.95 ± 0.28 | 70.34 ± 0.30 | 33.27 ± 0.17 | |
| ETF+raw radial | 24.14 ± 0.90 | 58.26 ± 0.53 | 28.37 ± 0.66 | |
| ETF+equal radius | 29.95 ± 0.28 | 70.34 ± 0.30 | 33.27 ± 0.17 | |
| ETF+no transport | 26.62 ± 0.33 | 67.72 ± 0.41 | 29.38 ± 0.43 | |
| Ours (RCR) | 32.24 ± 0.04 | 70.16 ± 0.16 | 35.46 ± 0.07 |
Table S5. Transport, radial reliability, and NC shrinkage are complementary.
Mixture Weight and Schedule
Table S6 shows that the fixed mixture is in a broad useful region rather than a hidden test-selected constant. Table S7 tests whether the deterministic stage schedule is necessary for the Seq-CIFAR-100 artifacts. The fixed-cap variant removes the stage ramp and is used by the main paper.
| Set | Class | Task | Route | |
| C10/200 | 0.00 | |||
| 0.25 | ||||
| 0.50 | ||||
| 0.75 | ||||
| Ours | ||||
| 1.00 | ||||
| C10/500 | 0.00 | |||
| 0.25 | ||||
| 0.50 | ||||
| 0.75 | ||||
| Ours | ||||
| 1.00 | ||||
| C100/200 | 0.00 | |||
| 0.25 | ||||
| 0.50 | ||||
| 0.75 | ||||
| Ours | ||||
| 1.00 | ||||
| C100/500 | 0.00 | |||
| 0.25 | ||||
| 0.50 | ||||
| 0.75 | ||||
| Ours | ||||
| 1.00 | ||||
| Tiny/200 | 0.00 | |||
| 0.25 | ||||
| 0.50 | ||||
| 0.75 | ||||
| Ours | ||||
| 1.00 | ||||
| Tiny/500 | 0.00 | |||
| 0.25 | ||||
| 0.50 | ||||
| 0.75 | ||||
| Ours | ||||
| 1.00 |
Table S6. Sensitivity to the ETF/NC-shrunk radial mixture weight.
| Buffer | Mixture rule | Class FAA | Class FF | Task FAA | Route | |
| 200 | Stage schedule | 2 | 49.55 | 19.31 | 87.49 | 51.21 |
| 200 | Ours: fixed cap | 2 | 49.49 | 16.43 | 87.39 | 51.18 |
| 500 | Stage schedule | 2 | 53.12 | 11.44 | 87.72 | 54.91 |
| 500 | Ours: fixed cap | 2 | 53.13 | 10.29 | 87.71 | 54.94 |
Table S7. Ablation of the RCR mixture schedule on Seq-CIFAR-100.
Note: Class, Task, and Route report final accuracies. The public method uses the deterministic fixed value , not a value selected on test data.
Table S7: Ablation of the RCR mixture schedule on Seq-CIFAR-100. The deployed fixed-cap rule uses at every task, while the stage schedule uses . Both variants use the NC-shrunk radial evidence and are evaluated on the same frozen artifacts; lower FF is better.
Note: The fixed-cap rule removes the deterministic stage ramp and is used by the reported Ours head. It reaches nearly the same final routing as the staged variant on these runs, while applying maximum radial evidence from the first task.
Analysis. The mixture ablations rule out a narrow schedule artifact. A broad interval of remains competitive, and removing the deterministic stage schedule does not remove the gain. This is why the main method uses the simpler fixed cap : it is tied to the number of classes per block rather than a tuned task index, while preserving the Class-IL and route improvements. Even the global value improves over ETF in every dataset/buffer cell in Table S6, so the gain is not created by a single test-selected mixture constant.
Radius and NC Retraction
Table S8 shows that the competitive radius exponent is not a narrow tuned constant; the default lies in a stable range. Table S9 isolates the NC-retraction clipping: the lower endpoint is the natural value 0, while the upper safeguard 0.95 avoids fully trusting transported centers after drift compensation.
| Class-IL | Task-IL | Route | Wrong-task | |
| 0.00 | 46.44 0.46 | 87.06 0.22 | 48.02 0.45 | 51.98 0.45 |
| 0.10 | 48.56 0.43 | 87.15 0.17 | 50.18 0.35 | 49.82 0.35 |
| Ours: 0.25 | 49.74 0.58 | 87.28 0.20 | 51.43 0.47 | 48.57 0.47 |
| 0.50 | 48.54 0.78 | 87.35 0.21 | 50.27 0.70 | 49.73 0.70 |
| 1.00 | 43.61 1.05 | 86.76 0.36 | 45.34 1.01 | 54.66 1.01 |
Table S8. Sensitivity to the competitive-radius exponent. Seq-CIFAR-100 buffer 200, final task. The default is stable: larger changes in the exponent preserve the routing gain, while the exact value is not selected on the test set.
| Dataset / Buffer | Setting | Class FAA | Class FF | |
| Seq-CIFAR-10 / 200 | No upper safeguard | 1.00 | 70.44 0.59 | 22.63 1.14 |
| Seq-CIFAR-10 / 200 | Ours | 0.95 | 70.36 0.50 | 22.85 1.19 |
| Seq-CIFAR-10 / 500 | No upper safeguard | 1.00 | 73.56 0.60 | 17.60 0.72 |
| Seq-CIFAR-10 / 500 | Ours | 0.95 | 73.68 0.72 | 17.73 0.72 |
| Seq-CIFAR-100 / 200 | No upper safeguard | 1.00 | 49.41 0.58 | 17.46 0.51 |
| Seq-CIFAR-100 / 200 | Ours | 0.95 | 49.74 0.58 | 16.27 0.37 |
| Seq-CIFAR-100 / 500 | No upper safeguard | 1.00 | 53.06 0.40 | 10.96 0.28 |
| Seq-CIFAR-100 / 500 | Ours | 0.95 | 53.27 0.34 | 10.18 0.31 |
| Seq-TinyImageNet / 200 | No upper safeguard | 1.00 | 30.57 0.58 | 26.46 0.82 |
| Seq-TinyImageNet / 200 | Ours | 0.95 | 31.42 0.65 | 23.84 0.94 |
| Seq-TinyImageNet / 500 | No upper safeguard | 1.00 | 31.45 0.12 | 25.12 0.20 |
| Seq-TinyImageNet / 500 | Ours | 0.95 | 32.24 0.04 | 22.47 0.10 |
Table S9. Upper clipping for NC retraction. We set and compare the deployed upper safeguard to the unconstrained endpoint . The lower clip is inactive on these runs; the useful safeguard is preventing full trust in transported centers.
Analysis. The radius and shrinkage tables show that RCR is not driven by a fragile radius constant. Moderate changes in keep the same qualitative behavior: larger class uncertainty reduces radial confidence, and the ETF component preserves global comparability. The NC retraction mainly stabilizes forgetting and late-stage routing by preventing unreliable transported centers from moving too far away from their neural-collapse vertices.
Replay Calibration Baselines
Replay-fitted calibrators are useful stress tests because they fit task or class biases from replay labels. They are not the same claim as RCR: RCR does not fit a task router from replay and does not add a task oracle. The comparison in Table S10 asks whether replay-side score calibration alone explains the gains. It does not: replay calibration can help in some TinyImageNet cells, but it is not consistently competitive with RCR and can damage CIFAR routing.
| Dataset/Buffer | Head | Class-IL ↑ | Task-IL ↑ | Route ↑ | Wrong-task ↓ |
| Seq-CIFAR-10 / 200 | ProNC ETF | 66.14±1.38 | 96.74±0.10 | 66.58±1.34 | 33.42±1.34 |
| Replay task-bias | 59.21±0.65 | 96.74±0.10 | 59.60±0.64 | 40.40±0.64 | |
| Replay class-bias | 59.36±0.62 | 96.72±0.12 | 59.75±0.60 | 40.25±0.60 | |
| Replay task affine | 58.19±1.35 | 96.74±0.10 | 58.61±1.32 | 41.39±1.32 | |
| Ours (RCR) | 70.36±0.50 | 96.67±0.08 | 70.82±0.46 | 29.18±0.46 | |
| Seq-CIFAR-10 / 500 | ProNC ETF | 72.16±1.30 | 96.82±0.12 | 72.71±1.35 | 27.29±1.35 |
| Replay task-bias | 67.89±0.91 | 96.82±0.12 | 68.35±0.97 | 31.65±0.97 | |
| Replay class-bias | 68.03±0.88 | 96.87±0.07 | 68.49±0.93 | 31.51±0.93 | |
| Replay task affine | 67.84±0.49 | 96.82±0.12 | 68.33±0.47 | 31.67±0.47 | |
| Ours (RCR) | 73.68±0.72 | 96.81±0.03 | 74.33±0.83 | 25.67±0.83 | |
| Seq-CIFAR-100 / 200 | ProNC ETF | 41.71±0.61 | 85.53±0.16 | 42.81±0.57 | 57.19±0.57 |
| Replay task-bias | 29.91±0.63 | 85.53±0.16 | 30.67±0.59 | 69.33±0.59 | |
| Replay class-bias | 29.36±0.65 | 84.86±0.06 | 30.13±0.70 | 69.87±0.70 | |
| Replay task affine | 27.54±2.57 | 85.53±0.16 | 28.35±2.56 | 71.65±2.56 | |
| Ours (RCR) | 49.74±0.58 | 87.28±0.20 | 51.43±0.47 | 48.57±0.47 | |
| Seq-CIFAR-100 / 500 | ProNC ETF | 48.65±0.55 | 85.54±0.19 | 50.14±0.42 | 49.86±0.42 |
| Replay task-bias | 36.14±0.85 | 85.54±0.19 | 37.06±0.91 | 62.94±0.91 | |
| Replay class-bias | 36.22±0.35 | 85.51±0.36 | 37.12±0.29 | 62.88±0.29 | |
| Replay task affine | 36.03±0.89 | 85.54±0.19 | 37.05±0.82 | 62.95±0.82 | |
| Ours (RCR) | 53.27±0.34 | 87.74±0.05 | 55.04±0.21 | 44.96±0.21 | |
| Seq-TinyImageNet / 200 | ProNC ETF | 25.77±0.41 | 68.30±0.62 | 28.58±0.31 | 71.42±0.31 |
| Replay task-bias | 28.48±0.11 | 68.30±0.62 | 31.43±0.11 | 68.57±0.11 | |
| Replay class-bias | 29.49±0.15 | 72.05±0.21 | 32.50±0.14 | 67.50±0.14 | |
| Replay task affine | 31.42±0.45 | 68.30±0.62 | 34.48±0.40 | 65.52±0.40 | |
| Ours (RCR) | 31.42±0.65 | 69.59±0.81 | 34.76±0.56 | 65.24±0.56 | |
| Seq-TinyImageNet / 500 | ProNC ETF | 26.61±0.32 | 67.71±0.39 | 29.37±0.42 | 70.63±0.42 |
| Replay task-bias | 29.80±0.16 | 67.71±0.39 | 32.59±0.17 | 67.41±0.17 | |
| Replay class-bias | 30.37±0.32 | 71.54±0.25 | 33.28±0.27 | 66.72±0.27 | |
| Replay task affine | 32.02±0.28 | 67.71±0.39 | 35.14±0.20 | 64.86±0.20 | |
| Ours (RCR) | 32.24±0.04 | 70.16±0.16 | 35.46±0.07 | 64.54±0.07 |
Table S10. Replay-fitted calibration does not explain RCR. Note: All calibration heads are fit only on replay-buffer scores and labels, then evaluated on held-out test features.
Analysis. Replay calibration is a useful negative control. If routing drift were only a global task-bias problem, replay-fitted offsets would match RCR. Instead, their behavior is inconsistent across datasets and memory budgets. This supports the main interpretation: the useful signal is not a learned task router, but the combination of transported class evidence and class-specific reliability under the same ETF decision space.
Mechanistic Diagnostics
The main paper argues that RCR acts when learned class measures deviate from the symmetric ETF geometry. The diagnostics below make this claim explicit through task-by-task curves, score-profile examples, stage-level NC proxies, route oracle bounds, and class-wise correlations.
Analysis
The diagnostic curves and block-score example visualize the same mechanism as the tables. RCR changes Class-IL and route accuracy together, while Task-IL is largely preserved. The block-score example makes this concrete: the sample is not a within-task class failure, because the correct class wins inside its own block; the error is caused by cross-block score comparability, which the reliability-calibrated score changes.
Analysis. Table S11 reports stage-level NC diagnostics and class-wise correlations, while Table S12 reports the true-route upper bound. Together they show that routing, not task-local discrimination, is the dominant remaining headroom. Finally, Table S13 and Figure S4 connect the effect to neural-collapse deviations at class level: classes with larger prototype-to-ETF deviation or larger competitive radius tend to have larger route error, which is exactly where radial reliability should matter.
Figure S4. Class-wise NC-deviation signals predict routing error. On Seq-CIFAR-100 with buffer 200, classes whose transported prototypes deviate further from their ETF vertices, or whose competitive radii are larger, tend to have larger held-out route error under the ProNC ETF head.
| Dataset/Buffer | Signal | Route error | Class |
| Seq-CIFAR-10 / 200 | |||
| NC1 disp. | |||
| Age | |||
| Seq-CIFAR-10 / 500 | |||
| NC1 disp. | |||
| Age | |||
| Seq-CIFAR-100 / 200 | |||
| NC1 disp. | |||
| Age | |||
| Seq-CIFAR-100 / 500 | |||
| NC1 disp. | |||
| Age | |||
| Seq-TinyImageNet / 200 | |||
| NC1 disp. | |||
| Age | |||
| Seq-TinyImageNet / 500 | |||
| NC1 disp. | |||
| Age |
Table S13. Class-wise NC-deviation correlations. Note: Pearson correlations are computed per seed across classes and then averaged.
References
- [1]Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample selection for online continual learning. In NeurIPS, 2019.arxiv.org/abs/1903.08671
- [2]Matteo Boschini, Lorenzo Bonicelli, Pietro Buzzega, Angelo Porrello, and Simone Calderara. Class-incremental continual learning into the extended der-verse. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5497–5512, 2023. doi: 10.1109/TPAMI.2022.3206549.DOI
- [3]Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. In NeurIPS, 2020.arxiv.org/abs/2004.07211
- [4]Hyuntak Cha, Jaeho Lee, and Jinwoo Shin. Co2L: Contrastive continual learning. In ICCV, 2021.
- [5]Arslan Chaudhry, Marc’Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a-GEM. In ICLR, 2019.arxiv.org/abs/1812.00420
- [6]Trung-Anh Dang, Vincent Nguyen, Ngoc-Son Vu, and Christel Vrain. Memory-efficient continual learning with neural collapse contrastive. In WACV, 2025.arxiv.org/abs/2412.02865
- [7]Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. PODNet: Pooled outputs distillation for small-tasks incremental learning. In ECCV, pages 86–102. Springer, 2020.arxiv.org/abs/2004.13513
- [8]Masih Eskandar, Tooba Imtiaz, Davin Hill, Zifeng Wang, and Jennifer Dy. Star: Stability-inducing weight perturbation for continual learning. In ICLR, 2025.arxiv.org/abs/2503.01595
- [9]Cong Fang, Hangfeng He, Qi Long, and Weijie J. Su. Exploring deep neural networks via layer-peeled model: Minority collapse in imbalanced training. Proceedings of the National Academy of Sciences, 118(43):e2103091118, 2021.DOI
- [10]Alex Gomez-Villa, Dipam Goswami, Kai Wang, Andrew D. Bagdanov, Bartlomiej Twardowski, and Joost van de Weijer. Exemplar-free continual representation learning via learnable drift compensation. In ECCV, pages 473–490. Springer-Verlag, 2024. doi: 10.1007/978-3-031-72667-5_27.
- [11]Dipam Goswami, Yuyang Liu, Bartlomiej Twardowski, and Joost van de Weijer. Fecam: Exploiting the heterogeneity of class distributions in exemplar-free continual learning. In NeurIPS, 2023.arxiv.org/abs/2309.14062
- [12]Zirui Guo, Quan Cheng, Da-Wei Zhou, and Lijun Zhang. Stable routing for mixture-of-experts in class-incremental learning, 2026. URL https://arxiv.org/abs/2605.17571.
- [13]X. Y. Han, Vardan Papyan, and David L. Donoho. Neural collapse under MSE loss: Proximity to and dynamics on the central path. In ICLR, 2022. URL https://openreview.net/forum?id=w1UbdvWH_R3.arxiv.org/abs/2106.02073
- [14]Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770–778, 2016.arxiv.org/abs/1512.03385
- [15]Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, pages 831–839, 2019.DOI
- [16]Gyuhak Kim, Changnan Xiao, Tatsuya Konishi, Zixuan Ke, and Bing Liu. A theoretical study on solving continual learning. In NeurIPS, 2022.arxiv.org/abs/2211.02633
- [17]Gyuhak Kim, Changnan Xiao, Tatsuya Konishi, and Bing Liu. Learnability and algorithm for continual learning. In ICML, 2023.arxiv.org/abs/2306.12646
- [18]James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114(13):3521–3526, 2017.DOI
- [19]Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009.
- [20]Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. Stanford University, CS231n course project report, 2015. URL https://cs231n.stanford.edu/reports/2015/pdfs/yle_project.pdf.
- [21]Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(12):2935–2947, 2018. doi: 10.1109/TPAMI.2017.2773081.DOI
- [22]Haowei Lin, Yijia Shao, Weinan Qian, Ningxin Pan, Yiduo Guo, and Bing Liu. Class incremental learning via likelihood ratio based task prediction. In ICLR, 2024.arxiv.org/abs/2309.15048
- [23]David Lopez-Paz and Marc’Aurelio Ranzato. Gradient episodic memory for continual learning. In NeurIPS, 2017.
- [24]Marc Masana, Xiailei Liu, Bartłomiej Twardowski, Mikel Menta, Andrew D. Bagdanov, and Joost Van De Weijer. Class-incremental learning: Survey and performance evaluation on image classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5513–5533, 2023. doi: 10.1109/TPAMI.2022.3213473.DOI
- [25]Michael McCloskey and Neal J. Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In Psychology of Learning and Motivation, volume 24, pages 109–165. Academic Press, 1989.DOI
- [26]Antoine Montmaur, Nicolas Larue, and Ngoc-Son Vu. Neural collapse inspired contrastive continual learning. In 35th British Machine Vision Conference 2024, BMVC 2024, Glasgow, UK, November 25–28, 2024. BMVA, 2024. URL https://papers.bmvc2024.org/0579.pdf.
- [27]Vardan Papyan, X. Y. Han, and David L. Donoho. Prevalence of neural collapse during the terminal phase of deep learning training. Proceedings of the National Academy of Sciences, 117(40):24652–24663, 2020. doi: 10.1073/pnas.2015509117.
- [28]Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. icarl: Incremental classifier and representation learning. In CVPR, 2017.arxiv.org/abs/1611.07725
- [29]Matthew Riemer, Ignacio Cases, Robert Ajemian, Miao Liu, Irina Rish, Yuhai Tu, and Gerald Tesauro. Learning to learn without forgetting by maximizing transfer and minimizing interference. In ICLR, 2019.arxiv.org/abs/1810.11910
- [30]David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy P. Lillicrap, and Gregory Wayne. Experience replay for continual learning. In NeurIPS, pages 350–360, 2019.
- [31]James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. CODA-Prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In CVPR, pages 11909–11919, 2023.arxiv.org/abs/2211.13218
- [32]Gjorgji Strezoski, Nanne van Noord, and Marcel Worring. Many task learning with task routing. In ICCV, 2019. doi: 10.1109/ICCV.2019.00146.DOI
- [33]Ruilin Tong, Yuhang Liu, Javen Qinfen Shi, and Dong Gong. Coreset selection via reducible loss in continual learning. In ICLR, 2025.
- [34]Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. FOSTER: Feature boosting and compression for class-incremental learning. In ECCV, pages 398–414. Springer, 2022.arxiv.org/abs/2204.04662
- [35]Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In CVPR, pages 5265–5274, 2018.arxiv.org/abs/1801.09414
- [36]Zheng Wang, Wanhao Yu, Li Yang, and Sen Lin. Rethinking continual learning with progressive neural collapse. In ICLR, 2026. URL https://openreview.net/forum?id=E3bBZ02Qcc.arxiv.org/abs/2505.24254
- [37]Yichen Wen, Zhiquan Tan, Kaipeng Zheng, Chuanlong Xie, and Weiran Huang. Provable contrastive continual learning. In ICML, pages 52819–52838, 2024.arxiv.org/abs/2405.18756
- [38]Liang Xie, Yibo Yang, Deng Cai, and Xiaofei He. Neural collapse inspired attraction–repulsion-balanced loss for imbalanced learning. Neurocomputing, 527:60–70, 2023.DOI
- [39]Yibo Yang, Haobo Yuan, Xiangtai Li, Zhouchen Lin, Philip Torr, and Dacheng Tao. Neural collapse inspired feature-classifier alignment for few-shot class incremental learning. In ICLR, 2023. URL https://openreview.net/forum?id=y5W8tpojhtJ.arxiv.org/abs/2302.03004
- [40]Yibo Yang, Haobo Yuan, Xiangtai Li, Jianlong Wu, Lefei Zhang, Zhouchen Lin, Philip Torr, Dacheng Tao, and Bernard Ghanem. Neural collapse terminus: A unified solution for class incremental learning and its variants, 2023. URL https://arxiv.org/abs/2308.01746.
- [41]Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In ICML, pages 3987–3995. PMLR, 2017.arxiv.org/abs/1703.04200
- [42]Chongyang Zhao, Mingsong Li, Haodong Lu, and Dong Gong. On token’s dilemma: Dynamic MoE with drift-aware token assignment for continual learning of large vision language models. In CVPR, 2026. URL https://arxiv.org/abs/2603.27481.






