Complexity-Stratified Curriculum Learning

• Complexity-Stratified Curriculum Learning is a strategy that organizes training by difficulty measures, exposing models gradually to harder data as proficiency increases.
• It integrates a difficulty measurer with a training scheduler to dynamically reweight examples, control pacing, and reduce gradient variance for effective learning.
• Empirical evidence shows that this approach significantly boosts performance across diverse fields, including computer vision, NLP, reinforcement learning, and quantum machine learning.

Complexity-Stratified Curriculum Learning denotes a family of training strategies in which data, tasks, or environments are exposed to a learner in an order induced by an explicit complexity or difficulty signal, typically progressing from easier strata to harder ones while controlling the pace of exposure. Within the general curriculum-learning framework, it is naturally expressed as a combination of a Difficulty Measurer and a Training Scheduler, and can be viewed as a sequence of data reweightings $Q_t(z)\propto W_t(z)P(z)$ that gradually increases diversity, entropy, and coverage until the target distribution is recovered (Wang et al., 2020). In this sense, complexity stratification is not a separate algorithmic family so much as a unifying formulation that covers manually predefined curricula, self-paced methods, transfer-teacher schemes, reinforcement-learning teachers, task-progress schedules, and more recent extensions to LLMs, offline reinforcement learning, graph learning, and quantum machine learning (Soviany et al., 2021).

1. Conceptual and mathematical formulation

At its most general, complexity-stratified training assigns each sample or task a scalar or vector difficulty signal and uses that signal to govern when the learner sees it. In the survey formulation, training is organized around a curriculum-weighted empirical risk,

$$L(\theta; t) = \sum_{i=1}^n w_i(t)\,\ell(y_i, f(x_i; \theta)) + R(\theta),$$

where $w_i(t)$ depends on a difficulty score $d_i$ and a pace parameter $\lambda(t)$. Binary inclusion uses

$w_i(t) = \mathbb{1}[d_i \le \lambda(t)],$

while soft weighting uses

$w_i(t) = g(d_i,\lambda(t)) \in [0,1],$

with $g$ decreasing with difficulty and increasing with pace (Wang et al., 2020).

This formulation is broad enough to cover instance-level curricula, task-level curricula, and mixed schedules. The same survey treats curriculum learning as a continuation-style procedure on non-convex objectives: easier or smoother training distributions are presented first, and harder or more diverse distributions are incorporated later (Wang et al., 2020). A closely related survey emphasizes that the same structure can be implemented through data selection, data weighting, progressive model or objective modification, or teacher–student control, and that stratification can be defined at the instance, task, or model level (Soviany et al., 2021).

In complexity-stratified settings, “complexity” is operational rather than purely semantic. It may refer to current loss, confidence margin, uncertainty, entropy, domain similarity, local density, noise level, structural simplicity, or any task-specific proxy that correlates with ease of learning. The unifying requirement is that lower-complexity strata be distinguishable from higher-complexity strata in a way that a scheduler can exploit (Wang et al., 2020).

2. Difficulty measures and stratification criteria

The central design decision is the difficulty signal itself. The survey literature groups such signals into loss-based measures, confidence or margin measures, uncertainty or entropy measures, task-specific signals, diversity or density measures, and source-reliability signals (Wang et al., 2020). Transfer-teacher curricula, for example, define difficulty from a pretrained model through quantities such as

$$m_i = p_{y_i}(x_i) - \max_{j\ne y_i} p_j(x_i), \qquad d_i=-m_i,$$

or

$d_i = H(p(x_i)) = -\sum_j p_j(x_i)\log p_j(x_i),$

or teacher loss $$L(\theta; t) = \sum_{i=1}^n w_i(t)\,\ell(y_i, f(x_i; \theta)) + R(\theta),$$0 (Wang et al., 2020).

Multi-criterion settings need not collapse to a single heuristic. The survey explicitly lists weighted aggregation,

$$L(\theta; t) = \sum_{i=1}^n w_i(t)\,\ell(y_i, f(x_i; \theta)) + R(\theta),$$1

as well as Pareto-based stratification over difficulty vectors $$L(\theta; t) = \sum_{i=1}^n w_i(t)\,\ell(y_i, f(x_i; \theta)) + R(\theta),$$2 (Wang et al., 2020). This becomes especially important when “hardness” has several independent sources, such as length and noise in NLP, or spatial interference and coordination dependence in multi-agent reinforcement learning.

Some later work makes the complexity definition itself the object of study. In parity learning, curriculum stages are defined by bias in product distributions rather than surface-level features. Cornacchia and Mossel show that under a biased product distribution, gradient signals reveal the parity support, whereas under the uniform distribution the target is orthogonal to lower-order features; complexity is therefore tied to the bias parameter rather than to superficial sample attributes (Cornacchia et al., 2023). In quantum curriculum learning, task similarity is defined through density ratios estimated with fidelity-based quantum kernels, and sample difficulty is further modulated by a dynamic threshold on current loss (Tran et al., 2024). In text-graph learning, difficulty is induced by graph complexity indices such as centrality, connectivity, and local bridges, together with text-complexity indices such as Coleman–Liau and sentence-length-related shallow features (Vakil et al., 2023).

A recurrent implication is that difficulty is domain-relative. The 2024 study on scoring functions for curriculum learning reports that common scoring functions are strongly dependent on the training setting, including randomness, that ensemble scoring can partly mitigate this dependence, and that robustness across random seeds positively correlates with curriculum-learning performance on CIFAR-10 (Rampp et al., 2024). This suggests that complexity stratification is only as reliable as the stability of the scoring rule that induces it.

3. Scheduling mechanisms and automatic curriculum construction

Once difficulty is defined, the scheduler determines how complexity is revealed. The survey lists inclusion thresholds, weighting functions, sampling distributions, and pace schedules such as linear, root-$$L(\theta; t) = \sum_{i=1}^n w_i(t)\,\ell(y_i, f(x_i; \theta)) + R(\theta),$$3, and geometric schedules (Wang et al., 2020). Representative pacing functions include

$$L(\theta; t) = \sum_{i=1}^n w_i(t)\,\ell(y_i, f(x_i; \theta)) + R(\theta),$$4

for linear pacing, exponential or geometric pacing, and quantile-based schedules that expose all samples satisfying $$L(\theta; t) = \sum_{i=1}^n w_i(t)\,\ell(y_i, f(x_i; \theta)) + R(\theta),$$5 (Wang et al., 2020).

Self-Paced Learning is the canonical automatic formulation. It jointly optimizes model parameters and latent sample weights:

$$L(\theta; t) = \sum_{i=1}^n w_i(t)\,\ell(y_i, f(x_i; \theta)) + R(\theta),$$6

With the hard regularizer

$$L(\theta; t) = \sum_{i=1}^n w_i(t)\,\ell(y_i, f(x_i; \theta)) + R(\theta),$$7

the optimal inclusion rule is

$$L(\theta; t) = \sum_{i=1}^n w_i(t)\,\ell(y_i, f(x_i; \theta)) + R(\theta),$$8

With a convex soft regularizer,

$$L(\theta; t) = \sum_{i=1}^n w_i(t)\,\ell(y_i, f(x_i; \theta)) + R(\theta),$$9

the optimal weights become

$w_i(t)$0

(Wang et al., 2020).

Teacher-driven curricula generalize this idea from sample selection to adaptive control. In teacher–student curriculum learning, a teacher observes task-wise progress and allocates future training to tasks with the largest magnitude slope, positive or negative, thereby handling both learning progress and forgetting (Matiisen et al., 2017). In the simple nonstationary-bandit formulation, the teacher updates an exponentially weighted reward estimate

$w_i(t)$1

and samples tasks via $w_i(t)$2-greedy or Boltzmann policies over $w_i(t)$3 (Matiisen et al., 2017). The related “Automated Curriculum Learning for Neural Networks” treats each task as a bandit arm and uses Exp3.S to optimize a stochastic syllabus from intrinsic rewards derived from prediction gain or complexity gain (Graves et al., 2017).

Reinforcement-learning curricula often expose complexity as a continuous environment parameter rather than a discrete task label. “Curriculum Learning with a Progression Function” defines a progression function $w_i(t)$4 that selects a complexity level $w_i(t)$5 from time and performance statistics, together with a mapping function $w_i(t)$6 that instantiates environments of that complexity (Bassich et al., 2020). A generic autonomous update is

$w_i(t)$7

and the paper emphasizes uniform sampling over all parameter settings up to the current frontier, rather than only at the frontier itself, in order to mitigate distribution shift (Bassich et al., 2020).

Recent masked-prediction work in offline RL replaces task order with mask order. CurrMask defines a pool of masking schemes using mask ratios $w_i(t)$8 and block sizes $w_i(t)$9, then uses EXP3 to adapt the masking distribution according to the decrease in a target masked-reconstruction loss (Tang et al., 2024). Complexity is expressed through temporal horizon, information sparsity, and compositionality: small blocks and low ratios are low complexity, while long-span masks induce planning-like dependencies (Tang et al., 2024).

4. Theoretical perspectives, mechanisms, and failure cases

Several theoretical accounts recur across the literature. The survey identifies optimization smoothing and continuation, variance reduction, implicit regularization, denoising, and distribution-shift control as the main explanations for why curriculum learning can help (Wang et al., 2020). Easier examples may exhibit lower gradient variance early in training, yielding faster SGD convergence; self-paced objectives connect to robust non-convex regularizers and MM formulations; and starting in high-confidence regions can minimize an upper bound of target risk under distribution shift (Wang et al., 2020).

The strongest formal positive result in the provided corpus comes from curriculum learning for parities. Cornacchia and Mossel prove that a suitable two-stage curriculum over product distributions can make learning $d_i$0-parities by SGD polynomial-time for small fully connected networks, whereas under the uniform distribution gradient-based learning faces a $d_i$1 lower bound (Cornacchia et al., 2023). They also prove a negative result for Hamming mixtures: curricula involving a bounded number of product distributions are not beneficial there, because Hamming-weight concentration prevents the staged distributions from resolving the target symmetry (Cornacchia et al., 2023).

A complementary RL theory is given by Rollin. There, a single difficult task is reformulated as a contextual multi-task problem whose contexts form a curriculum. Under Lipschitz reward variation across contexts and sufficiently small adjacent context gaps, policy transfer and roll-in from the previous task’s optimal visitation distribution let stochastic policy gradient skip the exponentially expensive bad-initialization phase. The resulting total complexity becomes polynomial, with the paper stating that curriculum and roll-in reduce complexity from exponential to polynomial in the state space (Li et al., 2022).

Not all evidence supports easy-to-hard sequencing. The survey explicitly notes that anti-curriculum or hard-example mining may be preferable on clean datasets, in settings where informative hard examples speed learning, or when emphasis on high-loss samples reduces class-imbalance bias; it also notes that some NMT settings favored anti-curriculum (Wang et al., 2020). This caution becomes sharper in recent LLM studies. A systematic post-training study on synthetic arithmetic and logical reasoning finds no robust advantage for difficulty-based sequencing over random sampling, whether under supervised fine-tuning or reinforcement learning, and reports that response lengths remain remarkably invariant across schedules (Mordig et al., 28 Mar 2026). A separate study on mathematical reasoning concludes that no curriculum strategy dominates universally and that forward versus reverse ordering depends jointly on model capability and task complexity (Jia et al., 21 Oct 2025).

These results do not collapse into a single verdict. A plausible implication is that complexity stratification is most helpful when difficulty is structurally aligned with transferable subskills, optimization conditioning, or exploration bottlenecks, and least helpful when the ordering signal is weakly related to the eventual generalization barrier.

5. Empirical manifestations across domains

Empirical work presents complexity-stratified curricula in several distinct forms. In computer vision and weakly supervised learning, the survey reports that curriculum methods improve convergence and generalization across classification, detection, segmentation, person re-identification, domain adaptation, and GAN training, and cites a 45.8% MAP improvement in multimedia event detection for self-paced learning with diversity (Wang et al., 2020). In unsupervised 3D medical image registration, curriculum by input blur anneals a Gaussian blur from $d_i$2 to $d_i$3 over the first 20,000 iterations and improves SLIVER test performance from Dice $d_i$4 and Jaccard $d_i$5 for the baseline 1-cascade VTN to Dice $d_i$6 and Jaccard $d_i$7, while preserving a better accuracy–speed trade-off than curriculum by smoothing (Burduja et al., 2021).

In natural language processing, competence-based machine translation curricula are reported to reduce training time by up to 70% and improve BLEU by up to 2.2 points (Wang et al., 2020). Curriculum demonstration selection for in-context learning stratifies demonstrations by metadata-defined difficulty and retrieves one example per stratum. On MATH, for example, Llama-2 7B improves from $d_i$8 with the base prompting setup to $d_i$9 with curriculum demonstration selection, while on Mercury code generation CodeLlama-13B improves Pass from $\lambda(t)$0 to $\lambda(t)$1 (Vu et al., 2024). Reverse curriculum generation by recursive dataset decomposition further extends the notion of complexity beyond ordering. On MATH-500, Qwen2.5-1.5B improves from $\lambda(t)$2 under full-dataset SFT to $\lambda(t)$3 with decomposition and to $\lambda(t)$4 with decomposition plus curriculum; on AIME ’25, Qwen3-4B-Base improves from $\lambda(t)$5 to $\lambda(t)$6 under decomposition plus curriculum (Zhao et al., 23 Feb 2026).

Reinforcement learning supplies some of the clearest task-level gains. In Minecraft Simon Says, a learning-progress curriculum combined with a dynamic exploration bonus increases the number of discovered items from $\lambda(t)$7 under uniform sampling without bonus to $\lambda(t)$8 under bidirectional learning progress plus dynamic bonus (Kanitscheider et al., 2021). Teacher–Student Curriculum Learning reports an order-of-magnitude speed-up over uniform sampling in a Minecraft maze and enables solving a maze that could not be solved at all when training directly on the final task (Matiisen et al., 2017). In offline RL, CurrMask achieves average reward $\lambda(t)$9 on zero-shot skill prompting compared with $w_i(t) = \mathbb{1}[d_i \le \lambda(t)],$0 for MaskDP, and average goal-conditioned planning distance $w_i(t) = \mathbb{1}[d_i \le \lambda(t)],$1 compared with $w_i(t) = \mathbb{1}[d_i \le \lambda(t)],$2 for MaskDP and $w_i(t) = \mathbb{1}[d_i \le \lambda(t)],$3 for Goal-GPT (Tang et al., 2024). In cooperative multi-agent reinforcement learning, a graph-based coordination complexity metric with $w_i(t) = \mathbb{1}[d_i \le \lambda(t)],$4 and $w_i(t) = \mathbb{1}[d_i \le \lambda(t)],$5 against empirical difficulty is used to order tasks; the resulting curriculum yields a reported 56x performance improvement in tight coordination tasks in MultiWalker (Ebadulla et al., 9 Jul 2025).

Other domains use analogous constructions. Quantum Curriculum Learning prioritizes auxiliary tasks by density ratios over quantum data and downweights hard or noisy quantum samples through a dynamic objective, improving convergence and robustness in unitary learning and quantum phase recognition (Tran et al., 2024). Complexity-Guided Curriculum Learning for Text Graphs combines graph-complexity and text-complexity indices with a spaced-repetition scheduler, improves node classification and link prediction, and is reported to use 39.2% fewer training examples than standard training for GTNN on Ogbn-Arxiv while improving accuracy from approximately $w_i(t) = \mathbb{1}[d_i \le \lambda(t)],$6 to approximately $w_i(t) = \mathbb{1}[d_i \le \lambda(t)],$7 (Vakil et al., 2023).

6. Relations to adjacent paradigms, controversies, and open problems

Complexity-stratified curriculum learning lies at the intersection of several neighboring ideas. The survey explicitly links it to transfer learning, meta-learning, continual learning, and active learning (Wang et al., 2020). Transfer-teacher curricula use pretrained models to define difficulty; meta-learning appears in example reweighting and “learning to teach”; continual-learning concerns enter through forgetting-aware schedules; and active-learning criteria such as uncertainty and diversity frequently reappear as curriculum signals (Wang et al., 2020).

The main controversy concerns whether ordering alone is causally important. The deductive-reasoning study argues that, under fixed compute budgets, structural data diversity and robust feedback mechanisms matter more than sequencing, and that random sampling suffices for the synthetic post-training tasks it considers (Mordig et al., 28 Mar 2026). The mathematical-reasoning study reaches a more conditional conclusion: task-aligned curricula affect final representations and generalization, whereas inner-state curricula modulate confidence and uncertainty, and medium-tier or decision-uncertain samples are often especially informative (Jia et al., 21 Oct 2025). The difficulty-definition study adds that there is no general advantage of curriculum learning over uniform sampling, but that data ordering matters and robustness of scoring functions positively correlates with curriculum performance on CIFAR-10 (Rampp et al., 2024).

Open problems identified in the survey remain largely intact: robust difficulty estimation for immature students, fairness and bias control so that minority classes or rare phenomena are not systematically delayed, scalable multi-task and multi-modal curricula, stronger theory for deep non-convex models, and adaptive teachers with lower computational overhead (Wang et al., 2020). The empirical record suggests that these issues are not secondary implementation details but constitutive parts of the subject. Complexity stratification is most coherent when the complexity signal is stable, domain-relevant, and matched to a scheduler that preserves diversity while exposing genuinely transferable structure. When those conditions fail, curriculum learning can degenerate into arbitrary ordering, or into a reweighting rule whose benefits are better explained by regularization than by “easy-to-hard” pedagogy.