Class-Incremental with Repetition (CIR)

• CIR is a continual learning scenario where previously seen classes reappear naturally, simulating realistic data streams.
• It employs stochastic stream generators and adaptive replay techniques to balance stability, plasticity, and dynamic class imbalance.
• By integrating unlabeled data and addressing natural repetition, CIR provides a more comprehensive benchmark than standard class-incremental learning.

Class-Incremental with Repetition (CIR) is a continual learning scenario in which new classes can appear over time and previously seen classes may reappear, simulating the recurrence seen in real-world data streams. In this family of scenarios, repetition is embedded in the definition of the stream, and repetition is a property of the environment and, unlike replay, cannot be controlled by the agent. CIR has been proposed as a corrective to the dominant class-incremental setting, which never allows revisiting previously seen classes and therefore neglects the role of repetition. Several recent formulations also attach abundant unlabeled data to the stream, making CIR a setting in which stability, plasticity, class recurrence, and selective use of unlabeled samples are studied jointly (Hemati et al., 2023, Cossu et al., 2021, Kim et al., 18 Aug 2025).

1. Formal setting and distinction from standard class-incremental learning

In standard class-incremental learning, the stream is split into a sequence of experiences, each introducing new classes, and the key property is the no repetition constraint

$$\mathcal{C}_i \cap \mathcal{C}_j = \emptyset, \quad \forall i \neq j .$$

Once a class is introduced, it never reappears in future experiences. CIR relaxes this constraint: previously encountered classes can reappear in future experiences, and each experience can contain a mix of new classes and classes already encountered in past experiences (Cossu et al., 2021).

This relaxation changes the interpretation of continual learning. The class-incremental scenario has been described as artificially exacerbating catastrophic forgetting, at the expense of other important objectives like forward transfer and computational efficiency. By contrast, repetition of previously encountered concepts occurs naturally in many real-world environments and contributes to softening the disruption of previous knowledge. Existing proposals such as the New Instances (NI) scenario, the New Instances and Classes (NIC) scenario, and CRIB were identified as examples in which repetition is integrated by design into the stream of incoming information (Cossu et al., 2021).

A central conceptual distinction in the CIR literature is the difference between repetition and replay. Replay is an intervention of the learning algorithm; repetition is a property of the environment. This distinction matters because a CIR stream may already contain natural revisitation of old concepts, while replay-based methods superimpose an additional, controllable rehearsal process (Hemati et al., 2023).

2. Stream construction, benchmark design, and scenario variants

A major contribution of the CIR literature is the construction of stream generators that move beyond fixed, disjoint class partitions. Two stochastic stream generators were introduced to produce a wide range of CIR streams starting from a single dataset and a few interpretable control parameters. The slot-based generator uses the number of experiences $N$ and the number of slots per experience $K$; it assigns each slot in each experience to a single class, ensures each sample is seen exactly once, and enables a transition from CI to CIR to DI by varying $K$. The sampling-based generator uses stream length $N$, experience size $S$, a distribution for first class occurrence $\mathcal{P}_f$, and class repetition probabilities $P_r$; it builds a binary occurrence matrix $T$, supports arbitrarily long streams, and can simulate imbalanced class repetition frequencies (Hemati et al., 2023).

A more general construction protocol was also described for CIR benchmark design. For a dataset $D$ and a specified number or sequence of experiences, the protocol defines which classes should appear in each experience and divides the original data so that each class’s samples are distributed equally across the experiences where that class appears. This was presented as a reproducible way to construct CIR scenarios for any classification dataset (Cossu et al., 2021).

Recent exemplar-free work further specialized CIR into stream families with partial repetition and mixing of tasks. Two scenarios were proposed: EFCIR-U 50/100 (Uniform Repetition) and EFCIR-B 50/100 (Biased Repetition). In EFCIR-U 50/100, after an initial task with 50 classes, the model trains on 99 additional small tasks with 2,000 samples per task, and all classes, seen or unseen, have a fixed 15% probability of appearing in any task. In EFCIR-B 50/100, the same structure is used, but per-class repetition probability is drawn from a Beta distribution with $N$0 and $N$1, introducing bias in repetition frequency (Tscheschner et al., 27 Feb 2025).

Stream family	Control parameters or structure	Main property
Slot-based generator	$N$2, $N$3	No instance repetition; transitions CI $N$4 CIR $N$5 DI
Sampling-based generator	$N$6, $N$7, $N$8, $N$9	Arbitrarily long; supports concept repetition and imbalance
EFCIR-U / EFCIR-B 50/100	Initial 50-class task, then 99 tasks of 2,000 samples	Partial repetition and mixing of tasks

3. Forgetting, natural rehearsal, and imbalance under repetition

The presence of repetition changes the empirical role of catastrophic forgetting. CIR studies report that as class repetition increases, even naive finetuning reduces forgetting and enables natural knowledge accumulation. In long streams generated by the sampling-based generator, repetition itself induces consolidation, and the gap between replay and naive strategies narrows as repetition increases. A related synthesis states that allowing even modest repetition can significantly reduce forgetting, and if repetition is abundant, continual learners’ accuracy can become close to that of joint training (Hemati et al., 2023, Cossu et al., 2021).

These findings do not imply that replay becomes irrelevant. Rather, they sharpen the distinction between generic forgetting mitigation and the specific problem of imbalance. CIR streams may be unbalanced: some classes are frequent, others rare, and replay buffers can inherit this skew. To address this, Frequency-Aware Replay (ER-FA) dynamically allocates replay quota inversely proportional to observed frequency,

$K$0

where $K$1 is the number of observations for class $K$2. The quota is normalized to fit the buffer size, so infrequent classes receive more slots (Hemati et al., 2023).

On unbalanced CIR streams, ER-FA was reported to improve both Average Classification Accuracy (ACA) and Missing Class Accuracy (MCA) on CIFAR-100 with fraction $K$3 (Hemati et al., 2023).

Strategy	ACA	MCA
ER-RS	51.1	16.7
ER-CB	55.3	34.5
ER-FA	57.8	44.7
DER-FA	63.4	51.2

CIR also exposes dynamic imbalance beyond static long-tailed learning. In rehearsal-based class-incremental learning, new classes are overrepresented while old classes are underrepresented, and this imbalance worsens dynamically across tasks as memory fills with more classes. In realistic scenarios such as CIR, some classes reappear, causing temporal, dynamic, and long-tailed imbalances that static approaches cannot handle well. This suggests that CIR is not only a softer alternative to strict class-incremental learning, but also a harder testbed for imbalance-aware learning rules (Chen et al., 2023).

4. Methodological families: rehearsal, classifier design, and exemplar-free adaptation

A canonical rehearsal-based approach aligned with CIR is iCaRL, which maintains exemplar sets $K$4 and classifies a query $K$5 by nearest-mean-of-exemplars: $K$6 When learning a new class, iCaRL selects a fixed number $K$7 of exemplars per class such that the total stored exemplars equal a memory budget $K$8, with $K$9 after learning $K$0 classes. It trains on both data for new classes and stored exemplars from previous classes, and for old classes it preserves previous outputs through distillation. In the CIR interpretation given in later summaries, iCaRL explicitly stores a small replay buffer and therefore operationalizes Class-Incremental with Repetition through direct repetition-based rehearsal (Rebuffi et al., 2016).

The dynamic imbalance of CIR motivated classifier redesign. The Dynamic Residual Classifier (DRC) builds on a residual classifier in which each task adds a branch layer $K$1, and residual fusion averages logits from current and previous branches: $K$2 To avoid linear model growth, DRC merges the previous merged branch and the last branch so that the model keeps a fixed size with two branches, one for “old” and one for “new”. It was presented as a drop-in replacement for the classifier head and reported to be compatible with major class-incremental pipelines such as MAF, MDT, and MEC, with state-of-the-art results on both the conventional CIL and the LT-CIL benchmarks (Chen et al., 2023).

CIR has also motivated exemplar-free methods that use repetition as a structural cue rather than as stored rehearsal. The Horde method dynamically adjusts an ensemble of self-reliant feature extractors and aligns them by exploiting class repetition. For each class $K$3, Horde maintains a prototype defined by the mean $K$4 and standard deviation $K$5 of embeddings, and projects pseudo-features by

$K$6

The paper reports that most classic prototype-based methods underperform or even degrade with repetition, whereas Horde outperforms all other exemplar-free methods decisively under both uniform and biased repetition. On CIFAR-100, the reported accuracies were 62.9% on CIL 50/10, 54.4% on EFCIR-U 50/100, and 54.3% on EFCIR-B 50/100 (Tscheschner et al., 27 Feb 2025).

5. CIR with unlabeled data

A recent branch of CIR research augments the stream with unlabeled data. In this formulation, classes can reappear over time, and training at each step can leverage both labeled data for current classes and easily accessible unlabeled external data. The unlabeled stream may include images from seen classes, unseen classes, or irrelevant classes which should be ignored. This makes CIR a setting in which selective exploitation of unlabeled data becomes part of the continual learning problem itself (Kim et al., 18 Aug 2025, Moraiti et al., 19 Mar 2025).

The CVPR 5th CLVISION Challenge operationalized this setting with a subset of ImageNet-1K containing 130 classes, of which 100 were learnable and 30 were held out or distractors. The stream consisted of 50 tasks, each with 500 labeled and 1,000 unlabeled samples, under an 8 GB GPU budget, a 600 minute limit, and a memory buffer capped at 200 exemplars. One solution introduced Multi-Level Knowledge Distillation (MLKD), which distills from multiple previous models at both feature and logit levels, and Dynamic Self-Supervised Loss (SSL), whose contribution decreases over tasks through $K$7. With a baseline of 19.54% final accuracy, the reported improvements were 23.82% for Baseline + Dynamic SSL, 39.82% for Baseline + MLKD, and 42.00% for MLKD + Dynamic SSL; the full method achieved 2nd place in the challenge (Kim et al., 18 Aug 2025).

Another challenge solution combined dual-stream Learning without Forgetting, feature consistency regularization, and prototype-based pseudo-labeling. At the start of each experience, class prototypes are computed as the mean feature embeddings of labeled samples, and an unlabeled sample receives a pseudo-label if its cosine similarity to the nearest prototype exceeds a confidence threshold $K$8; otherwise it is ignored. The reported average accuracies were 16.68% during the pre-selection phase and 21.19% during the final evaluation phase, compared with a baseline accuracy of 9.39%, yielding 4th place overall (Moraiti et al., 19 Mar 2025).

A plausible implication is that unlabeled-data CIR occupies an intermediate position between class-incremental learning, semi-supervised learning, and open-world recognition. In the cited challenge setting, stability and plasticity were not addressed solely by replay or distillation, but also by deciding which unlabeled observations should contribute to learning and which should be ignored (Kim et al., 18 Aug 2025).

6. Significance, misconceptions, and research directions

A recurring argument in the CIR literature is that class-incremental learning alone is not enough for continual learning evaluation. The excessive focus on the standard class-incremental scenario was said to be limiting because it overemphasizes catastrophic forgetting while underrepresenting forward transfer, computational efficiency, and adaptation under naturally recurring concepts. CIR was proposed precisely to restore these dimensions to the benchmark design (Cossu et al., 2021).

One common misconception is to treat repetition as equivalent to replay. CIR explicitly separates the two: repetition belongs to the environment, whereas replay is a strategy implemented by the learner. Another misconception is that methods that perform best on classical class-incremental benchmarks will necessarily perform best when classes repeat. Exemplar-free evaluations under EFCIR-U and EFCIR-B were used to argue that repetition fundamentally changes class-incremental dynamics and that strategies that perform best on classical CIL may fail, or even degrade, in realistic settings with class repetition (Tscheschner et al., 27 Feb 2025).

The current literature also indicates that CIR is not a single regime but a family of regimes. Some studies emphasize natural repetition and its effect on forgetting; others emphasize imbalance and long-tail effects; challenge settings emphasize unlabeled data, distractor classes, and small memory budgets. This suggests that “CIR” functions as an umbrella term for streams in which repetition is integrated by design, but the operational consequences depend strongly on how repetition, imbalance, and unlabeled data are instantiated (Hemati et al., 2023, Kim et al., 18 Aug 2025).

Taken together, these works present CIR as a more realistic and more discriminative testbed for continual learning systems. It preserves the class-incremental requirement that new concepts emerge over time, but it replaces the artificial ban on recurrence with a stream model in which old concepts may return, sometimes frequently, sometimes rarely, and sometimes mixed with unlabeled or irrelevant observations. Under that interpretation, CIR is not merely a variant of class-incremental learning; it is a reformulation of continual learning evaluation around the statistical structure of realistic streams (Cossu et al., 2021, Hemati et al., 2023).