Curriculum Continuity in Learning Systems

• Curriculum continuity is the systematic approach to preserving smooth progression in learning by managing transitions in difficulty, delivery, and structure across diverse domains.
• Adaptive pacing functions, homotopy methods, and replay strategies have been shown to enhance performance and reduce forgetting in both machine learning and educational settings.
• Empirical studies reveal that well-designed continuity mechanisms improve transfer, accuracy, and engagement, while balancing rigid schedules with adaptive strategies.

Curriculum continuity denotes the preservation of smooth progression in learning or teaching systems when difficulty, content, supervision, sequencing, delivery mode, or administrative structure changes over time. Across the literature, the term is not confined to a single formalism: in machine learning it appears as continuous pacing, homotopy-based loss interpolation, synthetic-to-real interpolation, and anti-forgetting replay schedules; in higher education it appears as seamless module sequencing, empirical pathway continuity through prerequisite structures, limited timetable perturbations across academic years, and feedback loops that preserve outcome alignment (Soviany et al., 2021, Pathak et al., 29 Jul 2025, Liang et al., 2024, Matiisen et al., 2017, Gupta et al., 2021, Paz, 5 Dec 2025, Meier et al., 19 Jun 2026, Derouich, 29 Oct 2025).

1. Conceptual scope and domain-specific meanings

The literature uses curriculum continuity to describe several related mechanisms that all reduce disruptive transitions. In continuous curriculum learning surveys, continuity is the gradual and smooth introduction of harder examples through a pacing function $p(t)$ or a monotonically increasing cutoff $\lambda(t)$, optionally replaced by soft sample weights $w_k(t)$ rather than hard inclusion rules (Soviany et al., 2021). In parameter-continuation methods for neural optimization, continuity is a smooth solution path in $(\theta,\lambda)$ space that connects an easy auxiliary objective to the target loss (Pathak et al., 29 Jul 2025). In diffusion-based data generation, it is a scalar guidance path from purely synthetic images to reconstructions of real ones (Liang et al., 2024). In continual learning, continuity is the ordering of tasks or replayed exemplars so as to maximize transfer and minimize forgetting (Singh et al., 2022, Tee et al., 2023, Matiisen et al., 2017). In university curriculum studies, continuity encompasses micromodule revision cycles, bridge arrangements during curriculum transitions, low-friction course-dependency topologies, semester-to-semester timetable stability, and continuous CLO–PLO realignment (Gupta et al., 2021, Karjanto et al., 2015, Paz, 5 Dec 2025, Meier et al., 19 Jun 2026, Derouich, 29 Oct 2025).

Domain	Continuity mechanism	Representative source
Neural network optimization	Homotopy in $(\theta,\lambda)$ space	(Pathak et al., 29 Jul 2025)
Data-level curriculum learning	Monotone pacing or soft weighting	(Soviany et al., 2021)
Synthetic data generation	Guidance parameter $\lambda \in [0,1]$ from synthetic to real	(Liang et al., 2024)
Continual learning	Task ordering, replay spacing, replay sequencing	(Singh et al., 2022, Tee et al., 2023, Matiisen et al., 2017)
Higher education design	Micromodule adaptation, bridge plans, graph continuity, perturbation control, feedback loops	(Gupta et al., 2021, Karjanto et al., 2015, Paz, 5 Dec 2025, Meier et al., 19 Jun 2026, Derouich, 29 Oct 2025)

This distribution of meanings suggests that curriculum continuity is less a single method than a family of continuity constraints on learning dynamics. The shared premise is that abrupt jumps in difficulty, representation, scheduling, or structural dependency are a source of instability, whether the learner is a neural network, a continual learner, or a student cohort.

2. Continuity as smooth optimization and difficulty pacing

A general formulation of continuous curriculum learning introduces a pacing function $\lambda(t)$ that determines which examples or losses are active at training step $t$. The survey literature lists linear, root, geometric, logarithmic, and sigmoidal schedules, and distinguishes between a discrete growing-batch curriculum and a continuous soft curriculum in which example weights evolve as $w_k(t)=\sigma(\lambda(t)-d_k)$ (Soviany et al., 2021). This framework formalizes continuity as gradual exposure rather than abrupt unlocking.

A more explicit mathematical treatment appears in "Principled Curriculum Learning using Parameter Continuation Methods" (Pathak et al., 29 Jul 2025). The training objective $L(\theta)$ is embedded into an easy-to-hard family by selecting an auxiliary loss $\lambda(t)$0 with known minimizer $\lambda(t)$1 and defining

$\lambda(t)$2

At $\lambda(t)$3, the system recovers the easy problem; at $\lambda(t)$4, it recovers the original critical-point condition. Under the stated $\lambda(t)$5 regularity and nonsingularity conditions, the Implicit Function Theorem guarantees a local $\lambda(t)$6 solution path $\lambda(t)$7. The paper’s central claim is that continuity in $\lambda(t)$8 is itself a curriculum, because the optimizer is warm-started in the basin of attraction of the next problem. To handle folds or bifurcations, the method reparameterizes the path by arc-length $\lambda(t)$9 and uses first-order Pseudo-arclength Continuation (PARC), a predictor–corrector scheme that combines a secant-step predictor with a constrained corrector on $w_k(t)$0 plus an orthogonality penalty. On down-sampled MNIST $w_k(t)$1, both NPC and PARC improved over vanilla ADAM in 4 out of 5 curriculum variants, and PARC gave the best path-tracking through non-monotonic solution manifolds.

"Curriculum DeepSDF" (Duan et al., 2020) implements continuity through two coupled curriculum axes: surface accuracy and sample difficulty. The model begins with coarse shape reconstruction by ignoring errors inside a tolerance band $w_k(t)$2, then tightens supervision as $w_k(t)$3 through

$w_k(t)$4

It simultaneously re-weights samples with a difficulty parameter $w_k(t)$5, up-weighting incorrectly signed and near-boundary points while down-weighting easy ones. The schedule runs for 2,000 epochs, progressively increasing the network from 5 to 8 fully-connected layers, stepping $w_k(t)$6 from $w_k(t)$7 to $w_k(t)$8, and increasing $w_k(t)$9 from $(\theta,\lambda)$0 to $(\theta,\lambda)$1. On ShapeNet chairs, sofas, tables, planes, and lamps, the average Chamfer Distance mean improved from $(\theta,\lambda)$2 for DeepSDF to $(\theta,\lambda)$3 for Curriculum DeepSDF, a $(\theta,\lambda)$4 reduction, while mesh accuracy mean improved from $(\theta,\lambda)$5 to $(\theta,\lambda)$6, a $(\theta,\lambda)$7 improvement.

In "Diffusion Curriculum: Synthetic-to-Real Data Curriculum via Image-Guided Diffusion" (Liang et al., 2024), continuity is the continuous scalar guidance parameter $(\theta,\lambda)$8 that interpolates between text-guided generation and real-image reconstruction. The method defines a guided noise prediction

$(\theta,\lambda)$9

and equivalently interprets the output as $(\theta,\lambda)$0. Hard samples are identified by classifier confidence $(\theta,\lambda)$1 and thresholded into a hard set $(\theta,\lambda)$2. The curriculum can be non-adaptive, using a monotone sequence $(\theta,\lambda)$3, or adaptive, selecting

$(\theta,\lambda)$4

On ImageNet-LT, the adaptive version improved all-class accuracy from $(\theta,\lambda)$5 to $(\theta,\lambda)$6 and tail-class accuracy from $(\theta,\lambda)$7 to $(\theta,\lambda)$8; on iWildCam, adaptive DisCL improved macro-accuracy from $(\theta,\lambda)$9 to $\lambda \in [0,1]$0.

A different instantiation appears in "Curriculum Multi-Task Self-Supervision Improves Lightweight Architectures for Onboard Satellite Hyperspectral Image Segmentation" (Carlesso et al., 16 Sep 2025). Here continuity is imposed not by changing task weights but by sorting training cubes according to a single difficulty score $\lambda \in [0,1]$1 derived from 3D gradient magnitudes and then exposing the encoder to nested datasets $\lambda \in [0,1]$2. Stage $\lambda \in [0,1]$3 uses the first

$\lambda \in [0,1]$4

sorted images and allocates

$\lambda \in [0,1]$5

epochs. This schedule never removes easy samples, so continuity is cumulative rather than replacement-based. The shared encoder is trained on masked image modeling, spatial jigsaw, and spectral jigsaw simultaneously. On Pavia University with 2D Justo, training from scratch yielded $\lambda \in [0,1]$6, MIM alone yielded $\lambda \in [0,1]$7, naive MTSSL yielded $\lambda \in [0,1]$8, and CMTSSL yielded $\lambda \in [0,1]$9.

Across these formulations, continuity can reside in the loss landscape, the data distribution, the supervision tolerance, or the exposure set. A plausible implication is that the most general invariant is not a particular schedule shape, but the requirement that the learner’s operating regime move along a controlled path rather than across discontinuous jumps.

3. Continuity as transfer management and anti-forgetting in continual learning

In continual learning, curriculum continuity is defined less by a scalar difficulty axis than by the preservation of knowledge while new tasks arrive. "Teacher-Student Curriculum Learning" formalizes this as a Teacher selecting subtasks for a Student from a finite set $\lambda(t)$0, with reward based on learning progress on the selected task (Matiisen et al., 2017). The setting is cast as a simple POMDP: the Student’s internal state $\lambda(t)$1 is unobserved, the Teacher’s action is the chosen subtask $\lambda(t)$2, the observation is a scalar score on that task, and the reward is the score change since the most recent prior training on the same task. Learning progress is estimated either by a linear-regression slope over recent observations or by a naive finite-difference estimate. Task selection becomes a non-stationary multi-armed bandit, and the use of $\lambda(t)$3 means that negative progress, interpreted as forgetting, increases the probability of revisiting the task. The paper argues that this avoids both stagnation on mastered tasks and jumps to tasks with zero learning signal. On decimal addition with LSTM, TSCL variants substantially outperformed uniform sampling and the best hand-tuned curricula; in the 1D case, Window + $\lambda(t)$4slope$\lambda(t)$5 reached mastery roughly $\lambda(t)$6 faster than the best manual schedule. In Minecraft navigation, uniform mixing and training only on the final task failed entirely on the hardest maze, while Window-based TSCL matched or slightly exceeded manual curriculum performance.

"Learning to Learn: How to Continuously Teach Humans and Machines" defines continuity in online class-incremental learning through two explicit criteria: maximize forward transfer and minimize forgetting (Singh et al., 2022). For a class sequence $\lambda(t)$7, the paper introduces average incremental accuracy $\lambda(t)$8, forgetting measure $\lambda(t)$9 for the first task, and the composite learning-effectiveness score

$t$0

Curricula are ranked by the Curriculum Designer (CD), which uses a distance confusion matrix $t$1 built from prototype feature vectors and cosine distance. The additive score $t$2 combines a low-variance first-class criterion, large early inter-class distances, and small late distances to previously seen classes. Across MNIST, FashionMNIST, CIFAR-10, and a human psychophysics experiment on a novel-object dataset, the paper reports that curriculum consistently influences learning outcomes, that CL algorithms tend to agree on which curricula are best, and that curricula beneficial to machines also tend to help humans. In the MTurk cohort, the top-5 curricula achieved average $t$3, whereas the worst-5 achieved $t$4.

"Integrating Curricula with Replays: Its Effects on Continual Learning" moves continuity into the replay mechanism itself (Tee et al., 2023). The study examines three dimensions: interleaved replay frequency, replay sequence, and exemplar selection. For interleaving, current-task data $t$5 and replay buffer $t$6 are partitioned into $t$7 blocks and alternated block by block; on ciFAIR-100 with a buffer size of 1200, forgetting dropped from $t$8 at $t$9 to $w_k(t)=\sigma(\lambda(t)-d_k)$0 at $w_k(t)=\sigma(\lambda(t)-d_k)$1, while average accuracy rose from $w_k(t)=\sigma(\lambda(t)-d_k)$2 to $w_k(t)=\sigma(\lambda(t)-d_k)$3. For replay order, easy-to-hard sequencing consistently outperformed hard-to-easy, with instance-level confidence-based difficulty slightly beating distance-based difficulty. For exemplar selection, uniform difficulty sampling outperformed selecting only the easiest or only the hardest examples. The combined curriculum transformed a baseline with average accuracy near $w_k(t)=\sigma(\lambda(t)-d_k)$4 and forgetting $w_k(t)=\sigma(\lambda(t)-d_k)$5 into a replay learner with average accuracy near $w_k(t)=\sigma(\lambda(t)-d_k)$6 and forgetting near $w_k(t)=\sigma(\lambda(t)-d_k)$7.

These works define continuity as a control principle over memory consolidation. The central object is not merely the next example’s difficulty, but the evolving relation between novelty and retention.

4. Modular, participatory, and transitional continuity in higher education

In higher education design, curriculum continuity is often treated as a property of delivery systems and governance procedures rather than only content order. The Montreal AI Ethics Institute’s post-pandemic university model describes a continuous learning system built from 10–15 minute micromodules with three metadata fields: Topic Tag, Relevance Score $w_k(t)=\sigma(\lambda(t)-d_k)$8, and Delivery Modality Flag (Gupta et al., 2021). Curriculum design is distributed across students and early-career learners, faculty and subject-matter experts, and industry and public-health partners. Micro-workshops follow a four-step process: Seed Module Release, Learner Exploration, Collaborative Synthesis Session, and Community Vote & Governance. Module relevance is updated by

$w_k(t)=\sigma(\lambda(t)-d_k)$9

where learner feedback and industry signal jointly determine future sequencing. Modules with $L(\theta)$0 are deprecated, but removal also requires both $L(\theta)$1 and $L(\theta)$2 to fall below $L(\theta)$3 and at least $L(\theta)$4 percent of faculty to register a low-confidence vote. The model emphasizes rapid “swap-in-swap-out” agility, but also identifies challenges in educator bandwidth, equity for under-resourced learners, accreditation readiness, and calibration of $L(\theta)$5, $L(\theta)$6, $L(\theta)$7, and $L(\theta)$8.

A concrete transitional case appears in the redesign of Foundation Engineering mathematics at the University of Nottingham Malaysia Campus (Karjanto et al., 2015). Over 2008/09 and 2009/10, the old 40-credit structure of three modules was replaced by five 10-credit units. The new system split the year-long 20-credit Foundation Mathematics module into Calculus 1 and Calculus 2 and inserted a standalone Mathematical Techniques unit. Continuity during the transition was maintained through an overlapped-delivery model: existing students completed the old module as planned, while the April 2009 intake began the new curriculum. Tutorial-based recap weeks and shared problem-solving workshops ensured that the cohort split by credit structure did not split pedagogically. The redesign also used Week 1 diagnostics to route weaker students into a 3 hr-lecture + 1 hr-tutorial stream and stronger students into the standard 2 hr + 1 hr format. Reported outcomes included lecturer agreement that semester-by-semester planning became more coherent, a reported 30 percent increase in self-rated confidence in problem solving among new-curriculum students by the end of Semester 1, and improvement in first-semester calculus pass rates from 72 percent under HG1FND to 84 percent under F40CA1.

A resilience-centered perspective is provided by the SUCRE serious-game workshop on engineering education (Waldeck et al., 17 Oct 2025). Here continuity is implied as the ability of a program to keep delivering its core learning objectives, or restore them rapidly, under VUCA trigger events. The workshop operates with Curriculum Identity Cards, Trigger Cards, and Impact Cards, and proceeds through situational awareness, vulnerability analysis, and adaptive capacity design. Engineering-specific risks include dependence on in-person lectures and hands-on labs, rigid accreditation-driven structures, reliance on specialized equipment and placements, and sensitivity to external stakeholders. The paper explicitly notes that it does not present a numeric continuity index. Its emphasis lies instead on mapping direct and cascading impacts and on adaptive strategies such as proactive trigger scanning, modular curriculum design, redundant delivery paths, flexible assessment schemes, cross-institutional collaboration, and periodic scenario drills.

Taken together, these studies treat continuity as an institutional capability. Sequencing remains important, but it is embedded in governance, infrastructure, transition planning, and crisis response.

5. Structural, temporal, and outcome-based continuity

A structural account of continuity appears in "The Topology of Hardship: Empirical Curriculum Graphs and Structural Bottlenecks in Engineering Degrees" (Paz, 5 Dec 2025). A curriculum graph $L(\theta)$9 is constructed empirically from student trajectories rather than official syllabi. For an ordered course pair $\lambda(t)$00, an edge $\lambda(t)$01 is inserted when $\lambda(t)$02 and $\lambda(t)$03. Continuity is then linked to low structural friction and low empirical hardship. The paper computes graph density

$\lambda(t)$04

longest path length $\lambda(t)$05, betweenness-based bottleneck centrality, blocking probability, time-to-progress, and a composite hardship index

$\lambda(t)$06

High-hardship curricula showed systematically higher cohort dropout rates, with Pearson $\lambda(t)$07, and older Civil Eng (1996) and Programmador Universitario (1996/2004) plans ranked about $\lambda(t)$08–$\lambda(t)$09 above the institutional mean. The paper’s interpretation is that deep chains and central bottlenecks turn small setbacks into long stalls or dropouts.

Temporal continuity across academic years is formalized in "A Multi-Objective Approach to Curriculum-Based Course Timetabling with Continuity Across Semesters" (Meier et al., 19 Jun 2026). A course offered annually has a preferred set of periods $\lambda(t)$10 inherited from the corresponding semester of the previous year; assigning it outside $\lambda(t)$11 is a perturbation. The continuity objective is

$\lambda(t)$12

and the model limits this by a bound $\lambda(t)$13. The paper separates lecturer and student objectives and explores their Pareto trade-off using the lexicographic $\lambda(t)$14-constraint method. On real-world instances from TUM Straubing, strict continuity severely restricted both stakeholder objectives, moderate perturbations yielded large gains, and diminishing returns appeared beyond roughly $\lambda(t)$15–$\lambda(t)$16 changes. For Instance 1, hypervolume increased from approximately $\lambda(t)$17 at $\lambda(t)$18 to $\lambda(t)$19 at $\lambda(t)$20.

Outcome-based continuity is addressed in "Ensuring Outcome-Based Curriculum Coherence through Systematic CLO-PLO Alignment and Feedback Loops" (Derouich, 29 Oct 2025). The framework maps questions to CLOs via weights $\lambda(t)$21, CLOs to PLOs via weights $\lambda(t)$22, and then tracks intended versus delivered emphasis through teaching units and student assessment components. Core indicators include

$\lambda(t)$23

with analogous PLO-level ratios. The framework uses an acceptance band of $\lambda(t)$24–$\lambda(t)$25; values below $\lambda(t)$26 indicate under-delivery or under-assessment, while values above $\lambda(t)$27 indicate over-delivery or over-assessment. The proposed feedback loop adjusts teaching-unit allocation, assessment design, and CLO–PLO mappings across successive offerings.

These three lines of work show that continuity can be represented as graph connectivity, timetable perturbation control, or alignment stability across teaching and assessment. The underlying object changes, but the diagnostic pattern remains the same: continuity fails where dependencies, timing, or outcome mappings become discontinuous.

6. Empirical regularities, limits, and recurring points of tension

Several empirical regularities recur across the literature. Continuity mechanisms frequently improve downstream quality measures relative to discontinuous or non-adaptive baselines. Parameter continuation improved generalization over ADAM in most reported MNIST curriculum variants (Pathak et al., 29 Jul 2025). Curriculum DeepSDF reduced average Chamfer Distance by $\lambda(t)$28 and improved mesh accuracy by $\lambda(t)$29 (Duan et al., 2020). DisCL improved macro-accuracy on both long-tail and low-quality settings, including a rise from $\lambda(t)$30 to $\lambda(t)$31 on iWildCam and from $\lambda(t)$32 to $\lambda(t)$33 on ImageNet-LT (Liang et al., 2024). CMTSSL improved lightweight hyperspectral segmentation relative to both scratch training and naive multi-task self-supervision (Carlesso et al., 16 Sep 2025). In continual learning, TSCL, Curriculum Designer rankings, and replay curricula all showed that ordered exposure can improve transfer and reduce forgetting (Matiisen et al., 2017, Singh et al., 2022, Tee et al., 2023). In academic settings, bridge mechanisms, structural redesign, and alignment feedback loops were associated with smoother progression, improved planning coherence, and, in some cases, improved pass rates or reduced structural hardship (Karjanto et al., 2015, Paz, 5 Dec 2025, Derouich, 29 Oct 2025).

The literature also identifies clear limits. Continuity mechanisms are not uniformly beneficial in every instantiation: in the one-layer classifier reported with PARC, the h-Brightness variant yielded test accuracy $\lambda(t)$34, below the ADAM baseline at $\lambda(t)$35 (Pathak et al., 29 Jul 2025). Surveyed pacing methods introduce additional hyper-parameters such as $\lambda(t)$36, $\lambda(t)$37, and $\lambda(t)$38, and poorly chosen difficulty measures can bias the curriculum and degrade performance (Soviany et al., 2021). The MAIEI model explicitly warns about educator bandwidth, equity constraints, and accreditation readiness (Gupta et al., 2021). The timetabling formulation shows that continuity can directly conflict with lecturer and student objectives, so preserving prior-year slots too rigidly may lower timetable quality (Meier et al., 19 Jun 2026). The SUCRE workshop, by not providing a numeric continuity index, also indicates that some educational continuity problems are presently handled through structured deliberation rather than closed-form optimization (Waldeck et al., 17 Oct 2025).

A recurring point of tension concerns whether continuity should be rigid or adaptive. The evidence presented here favors adaptive continuity: DisCL selects guidance levels by immediate validation lift, TSCL revisits tasks when slopes become negative, micromodule systems revise relevance scores weekly, Nottingham’s transition model allowed dynamic regrouping into the 2+2 stream, and semester-to-semester timetabling treats perturbation limits as a decision parameter rather than an absolute prohibition (Liang et al., 2024, Matiisen et al., 2017, Gupta et al., 2021, Karjanto et al., 2015, Meier et al., 19 Jun 2026). This suggests that continuity is not equivalent to immobility. It is the controlled preservation of progression under change.

Curriculum continuity therefore functions as a unifying principle across optimization, continual learning, educational design, curriculum analytics, and administrative planning. In each case, it is the attempt to preserve a traversable path—through parameter space, data space, task space, course networks, academic calendars, or outcome maps—so that learning systems can change without losing coherence.