Dual-Learner Framework Overview

• Dual-Learner Framework is an architectural approach featuring two concurrent learners, one for rapid adaptation and another for stable knowledge consolidation.
• It leverages cumulative parameter averaging to merge past and present learning, effectively addressing exemplar-free incremental learning challenges on benchmarks like CIFAR-100.
• The framework’s dual roles extend to diverse fields such as reinforcement learning and treatment effect estimation, demonstrating versatile application potential.

Searching arXiv for the cited DLCPA paper and closely related "dual-learner" formulations to ground the article. Search query: arXiv (Sun et al., 2023) DLCPA dual learner cumulative parameter averaging exemplar-free incremental learning A dual-learner framework is an architectural pattern in which two learners are maintained concurrently and assigned complementary roles. In exemplar-free incremental learning, the formulation is instantiated by the Dual-Learner framework with Cumulative Parameter Averaging (DLCPA), which separates rapid adaptation from long-term retention through a plastic learner, a stable learner, and task-specific classifiers (Sun et al., 2023). The same designation also appears in other research areas: Dual Policy Distillation maintains two peer policies that learn from each other in reinforcement learning, the X-Learner and Robust X-Learner organize estimation through paired response and treatment-effect learners in heterogeneous treatment effect estimation, and the Prototype Expansion Network uses an Intrinsic Learner and a Diffusion Learner to construct complementary prototypes for few-shot 3D segmentation (Lai et al., 2020, Uehara, 21 Jan 2026, Zhao et al., 16 Sep 2025). This suggests that the term denotes a recurrent two-model coordination scheme rather than a single canonical algorithm.

1. Plasticity–stability as the canonical continual-learning setting

In incremental learning, the central difficulty is the dilemma between plasticity and stability, especially in the exemplar-free scenario where accessing old-task samples is strictly prohibited during the learning of a new task (Sun et al., 2023). A straightforward solution is Single Task Learning, in which an independent model is learned and stored for each task. Although Single Task Learning exhibits linear growth in model storage with the number of tasks, the key empirical observation behind DLCPA is that averaging these model parameters can potentially preserve knowledge across all tasks.

DLCPA was proposed from that observation. Its stated objective is to make exemplar-free incremental learning both plastic and stable by disentangling the acquisition of new-task knowledge from the accumulation of previously learned knowledge. In this formulation, the plastic learner affords rapid adaptation, while the stable learner accumulates consolidated knowledge. Task-specific classifiers mediate prediction and close the training loop by aligning decision layers to the feature space of the stable learner.

The framework is evaluated in both Task-IL, where task identity is known at test time, and Class-IL, where no task identity is provided. Within that setting, DLCPA is reported to outperform several state-of-the-art exemplar-free baselines on CIFAR-100 and Tiny-ImageNet (Sun et al., 2023).

2. Core architecture and notation

DLCPA contains three module types: a plastic learner $\theta_P$, a stable learner $\theta_S$, and task-specific classifiers $\{\phi_1,\dots,\phi_T\}$ (Sun et al., 2023). The plastic learner is a standard deep network such as ResNet-18 with ReLU activations. When learning task $t$, its parameters are denoted $\theta_P^{(t)}$, and it is free to adapt aggressively to new data. The stable learner has the same architecture but separate weights $\theta_S^{(t)}$; it serves as a running “mean” of past plastic learners and therefore preserves consolidated knowledge. For each task $t$, the classifier $\phi_t$ is a linear layer with one column per class in $Y_t$.

The inference path depends on the evaluation protocol. In Task-IL, the stable learner output $f(x\mid\theta_S)$ is fed into the appropriate $\theta_S$0. In Class-IL, the stable learner output is fed into the concatenation $\theta_S$1. During training of task $\theta_S$2, the plastic learner adapts to $\theta_S$3 under self-supervised and supervised losses; its updated parameters are then softly averaged into the stable learner; finally $\theta_S$4 is retrained with $\theta_S$5 frozen so that its prototypes align with features $\theta_S$6.

The principal parameter definitions are

$\theta_S$7

$\theta_S$8

with initialization $\theta_S$9, either random or pretrained (Sun et al., 2023).

3. Cumulative parameter averaging and cyclic updates

The defining mechanism of DLCPA is cumulative parameter averaging. After learning task $\{\phi_1,\dots,\phi_T\}$0, the stable learner is defined as the exact average over all plastic learners: $\{\phi_1,\dots,\phi_T\}$1 This can be implemented incrementally as

$\{\phi_1,\dots,\phi_T\}$2

The paper contrasts this rule with two ablated alternatives: exponential moving average,

$\{\phi_1,\dots,\phi_T\}$3

which was found inferior, and hard copy,

$\{\phi_1,\dots,\phi_T\}$4

which forgets too much (Sun et al., 2023).

Training for each new task proceeds in three repeated steps until convergence. First, the plastic learner is updated while $\{\phi_1,\dots,\phi_T\}$5 and $\{\phi_1,\dots,\phi_T\}$6 are frozen. The objective is

$\{\phi_1,\dots,\phi_T\}$7

where $\{\phi_1,\dots,\phi_T\}$8 can be BYOL, SimCLR, or MoCoV2, and $\{\phi_1,\dots,\phi_T\}$9 is standard cross-entropy. The hyperparameter $t$0 controls the balance, and $t$1 is reported to work well in practice (Sun et al., 2023).

Second, the stable learner is updated by cumulative averaging: $t$2 Third, with $t$3 frozen, the classifier $t$4 is retrained by minimizing

$t$5

In the paper’s interpretation, this final step re-aligns class prototypes to the stable feature space (Sun et al., 2023).

4. Empirical profile of DLCPA

DLCPA is evaluated on CIFAR-100, with 100 classes split into 5, 10, or 20 tasks, and on Tiny-ImageNet, with 200 classes split into 10 or 25 tasks (Sun et al., 2023). The reported metrics are final overall classification accuracy after $t$6 tasks (ACC%) and backward transfer (BWT%), defined as the average drop on old tasks.

Selected Task-IL results are reported for 10-split CIFAR-100 and 25-split Tiny-ImageNet:

Method	ACC	BWT
VDFD	83.30	-1.27
DLCPA+BYOL	83.15	-0.04
DLCPA+SimCLR	82.03	-0.99
DLCPA+MoCoV2	81.15	-0.34

For Class-IL on the 10-split setting, the reported accuracies are VDFD at 38.38, DLCPA+MoCoV2 at 40.11, and DLCPA+SimCLR at 40.82 (Sun et al., 2023). The BYOL version is additionally noted to yield slight positive BWT on Tiny-ImageNet.

Ablation results isolate the contribution of each component. On 10-split CIFAR-100 and Tiny-ImageNet with the MoCoV2 version, removing self-supervision reduces Task-IL ACC from 81.15 to approximately 64.7. Replacing cumulative averaging with exponential moving average or hard copy reduces Task-IL ACC to approximately 75. Training $t$7 on plastic features rather than stable features also reduces Task-IL ACC to approximately 64.7. Sensitivity analysis shows a stable peak around $t$8 for both datasets (Sun et al., 2023).

These results support the paper’s claim that the three ingredients—self-supervision, cumulative averaging, and classifier alignment to the stable learner—are jointly structural rather than auxiliary.

5. Why averaging and dual roles can work

The motivation for cumulative parameter averaging is not presented as a generic ensemble heuristic, but as a taskwise consolidation rule derived from empirical and theoretical observations (Sun et al., 2023). Empirical linear-probe studies show that an averaged extractor retains discrimination on all tasks. The paper also gives a theoretical Lipschitz argument bounding loss degradation by the “flatness” of task losses, and it states that flattening is aided by self-supervision.

Within that interpretation, the plastic learner is optimized to move rapidly in response to the current task, while the stable learner changes only through the averaging rule. The separation creates asymmetric roles: one learner is allowed to specialize, and the other is constrained to accumulate. Task-specific classifiers then provide indirect communication between $t$9 and $\theta_P^{(t)}$0 by being retrained against the stable feature space. A plausible implication is that DLCPA addresses plasticity and stability less by direct regularization of a single network than by role separation plus controlled parameter transfer.

The ablation evidence is consistent with that interpretation. Exponential moving average is found inferior, which indicates that the exact cumulative mean over past plastic learners is not interchangeable with a recency-biased smoother. Hard copy forgets too much, which is consistent with the stable learner’s role as a consolidated representation rather than a mirror of the newest model. Likewise, classifier training on plastic features rather than stable features sharply degrades performance, indicating that the prediction layer must be anchored to the accumulated feature space rather than the transient one (Sun et al., 2023).

6. Broader meanings of “dual-learner framework” across fields

Outside exemplar-free incremental learning, the same phrase describes related two-model coordination schemes with different coupling mechanisms. In Dual Policy Distillation, two parametric policies, $\theta_P^{(t)}$1 and $\theta_P^{(t)}$2, interact with the same MDP, explore concurrently, and alternate between a reinforcement-learning update and a distillation update (Lai et al., 2020). The key theoretical construct is a hypothetical hybrid policy that chooses, state by state, the action distribution of whichever policy has higher state-value. The practical surrogate distills only on states where the peer is advantageous, using weighted losses

$\theta_P^{(t)}$3

with a symmetric loss for $\theta_P^{(t)}$4. On HalfCheetah and Walker2d, DPD-DDPG is reported to achieve more than 15% higher maximum return at $\theta_P^{(t)}$5 steps than DDPG at the same step count, without teacher models or expert trajectories.

In heterogeneous treatment-effect estimation, the X-Learner is described as a three-step dual-learner framework built from two base-learner estimates $\theta_P^{(t)}$6 and $\theta_P^{(t)}$7, followed by cross-imputation and aggregation (Uehara, 21 Jan 2026). For treated units,

$\theta_P^{(t)}$8

and for control units,

$\theta_P^{(t)}$9

The Robust X-Learner replaces MSE with a redescending $\theta_S^{(t)}$0-divergence objective,

$\theta_S^{(t)}$1

to prevent “Outlier Smearing.” On a semi-synthetic Criteo Uplift dataset, it reduces Core-PEHE from $\theta_S^{(t)}$2 for the standard X-Learner to $\theta_S^{(t)}$3 at $\theta_S^{(t)}$4, a 98.6% reduction in error.

In few-shot 3D point cloud semantic segmentation, the Prototype Expansion Network uses a dual-stream learner architecture consisting of an Intrinsic Learner and a Diffusion Learner (Zhao et al., 16 Sep 2025). The Intrinsic Learner is a fully supervised DGCNN backbone, while the Diffusion Learner is a frozen, pre-trained conditional encoder from a 3D diffusion model. These learners generate intrinsic prototypes $\theta_S^{(t)}$5 and diffusion prototypes $\theta_S^{(t)}$6, which are aligned by a Prototype Assimilation Module and fused as

$\theta_S^{(t)}$7

A Prototype Calibration Mechanism regularizes the fused prototype bank. On S3DIS 2-way 1-shot, PENet reports 70.04% mIoU versus 66.41% for Seg-PN; on ScanNet 2-way 1-shot, it reports 68.86% versus 63.74% (Zhao et al., 16 Sep 2025).

Taken together, these cases show that a dual-learner framework can mean cumulative parameter averaging, advantageous peer distillation, robust cross-imputation, or cross-guided prototype fusion. What they share is explicit specialization between two learners and an explicit transfer rule between them.

7. Limitations and open problems

The limitations of DLCPA are explicit. It relies on clearly demarcated tasks, so extension to fully task-agnostic streams remains open (Sun et al., 2023). It also assumes tasks are roughly equal in importance, which leaves adaptive weighting or task prioritization as unresolved directions. The paper further identifies dynamic task grouping and combinations with exemplar replay or generative replay as possible extensions.

Related dual-learner frameworks expose other constraints. Dual Policy Distillation assumes sufficiently accurate value estimates for advantage-weighting; in very large MDPs or with biased critics, noisy $\theta_S^{(t)}$8 may mislead distillation (Lai et al., 2020). It also uses exactly two learners, and extending to $\theta_S^{(t)}$9 learners raises unresolved questions of pairwise distillation and consensus formation. The Robust X-Learner addresses heavy tails and imbalance through a non-convex objective, which in turn requires a Proxy Hessian strategy grounded in Majorization-Minimization principles to stabilize boosting updates (Uehara, 21 Jan 2026). PENet addresses semantic drift by adding a Prototype Calibration Mechanism, implying that dual-source prototype fusion alone is insufficiently constrained in few-shot settings (Zhao et al., 16 Sep 2025).

A plausible synthesis is that dual-learner frameworks do not eliminate the need for auxiliary structure; rather, they relocate it. In DLCPA, the auxiliary structure is cumulative averaging and classifier realignment. In DPD, it is advantage estimation and weighted distillation. In RX-Learner, it is robust loss design and MM-based optimization. In PENet, it is prototype assimilation and calibration. The recurring pattern is therefore not merely “two learners,” but two learners plus a carefully specified mechanism for transferring, filtering, or reconciling information between them.