---
kind: concept
slug: curriculum-learning
title: Curriculum learning
summary: >-
  Schedules training so that simpler examples or easier targets come first. Most order-based variants in a
  31-team small-data shared task did not beat a random order. Other variants schedule task difficulty, a
  pseudo-label threshold, a loss weight, an input source or a staged training pipeline instead. Test against a fixed-schedule baseline before adopting one.
tags: [training, data-ordering, pretraining, semi-supervised, pseudo-labeling, reinforcement-learning, robotics, robustness]
papers: ["2504.08165", "1809.03036", "2110.08263", "2103.09504", "2408.07295", "2610.17635", "2003.04960", "2502.05855", "2207.04672", "2508.11849", "2112.08304", "2112.05504"]
related: [neural-radiance-field, level-of-detail-rendering, dataset-curation, data-augmentation, knowledge-distillation, active-learning, language-model-pretraining, semi-supervised-learning, self-training, consistency-regularization, domain-randomization, sim-to-real-transfer, noisy-labels, partial-label-learning, transfer-learning, experience-replay, self-play, hierarchical-reinforcement-learning, vision-language-action-model, imitation-learning, adversarial-training, adversarial-robustness, minimax-optimization]
status: draft
updated_at: 2026-11-01
---

## Overview

Curriculum learning presents training examples in an order chosen by some complexity measure, so that
simpler examples come first. The BabyLM organisers defined it as sorting training steps by one or more
complexity metrics [arxiv:2504.08165 §7.4]. The shared invariant of the data-order variants is that the set of
examples stays fixed and only their order or schedule changes. Some variants also change the data itself, which
matters when comparing results [arxiv:2504.08165 app. F].

Twelve cards carry this concept, and they do not all change the same thing. Eleven list it under `concepts:`; the
partial multi-label card carries it only as a tag. The BabyLM Challenge findings cover
data order, where 13 of 31 teams (41.9%) tried some form of curriculum [arxiv:2504.08165 §7.4]. The paper treats
curriculum as its most common approach, and reports that most attempts did not give consistent gains over naive
ordering [arxiv:2504.08165 §7.4]. FlexMatch schedules a per-class confidence threshold for pseudo labels, not the
order of examples [arxiv:2110.08263 §3.1]. The human-motion model schedules a loss weight
[arxiv:1809.03036 §3.4]. The humanoid controller moves through three training stages of rising difficulty, which
change the task distribution rather than a list of examples [arxiv:2408.07295 §4.4]. The quadruped locomotion policy ramps obstacle density from sparse to dense over training
[arxiv:2508.11849 §4.4]. PredRNN's reverse scheduled
sampling changes how often the encoder sees real frames [arxiv:2103.09504 §4.4]. The partial multi-label method
recomputes a per-entry loss weight every epoch, and its paper calls that weighted loss a curriculum loss
[arxiv:2610.17635 eq. 1]. The DexVLA robot policy lists this concept, though its card describes three
training stages that change the data and the trainable parts rather than the order of examples
[arxiv:2502.05855 §3.2]. The reinforcement-learning survey gives the most general frame: a curriculum is a directed
acyclic graph over experience, and building one has three parts, which are task generation, sequencing and
transfer [arxiv:2003.04960 §3.1] [arxiv:2003.04960 §3.2]. It runs no experiments, so it maps methods rather than
reporting results. The multilingual translation model NLLB-200 adds groups of language pairs at fixed update
counts, so its schedule changes the data mix and not only the order [arxiv:2207.04672 §8.2.3]. The variants below
are grouped by what they change.

## Variants

### Complexity-ranked order

Examples are sorted by a score and presented in that order. The scores the paper lists include sentence
length, lexical frequency, syntactic complexity, and surprisal, and whole datasets were sorted by difficulty
[arxiv:2504.08165 §7.4]. Card: [Findings of the BabyLM Challenge](/papers/2504.08165). Several of these
submissions found no consistent gain over a random order [arxiv:2504.08165 app. F].

### Staged or two-phase schedule

Training begins with one ordering and then switches. One submission ranked sentences by surprisal, trained on
that order, and then trained further on a random-order pass. The random-order phase is what beat its baseline
[arxiv:2504.08165 app. F]; see the surprisal entry under Model-selected examples. Another, the Contextualizer,
sorted its corpora from easy to hard at the start and end of training and used a noisier arrangement in the
middle [arxiv:2504.08165 app. F]. It also built extra samples by combining chunks of text from different
contexts. The paper itself calls this a form of data augmentation [arxiv:2504.08165 §7.4], so its result
cannot be credited to the ordering alone [arxiv:2504.08165 §7.2]. Card:
[Findings of the BabyLM Challenge](/papers/2504.08165).

### Model-selected examples

A model scores each example and the schedule is built from those scores. One submission ranked sentences by
average surprisal from an ensemble of LSTM teachers. It did not beat a random baseline alone, but improved when
followed by random-order training [arxiv:2504.08165 app. F]. Another selected training examples by n-gram
surprisal and gained on GLUE but lost on MSGS [arxiv:2504.08165 app. F]. Card:
[Findings of the BabyLM Challenge](/papers/2504.08165).

### Growing vocabulary or objective

The vocabulary grows during training, or the training objective becomes harder over time
[arxiv:2504.08165 §7.4]. Card: [Findings of the BabyLM Challenge](/papers/2504.08165).

### Language-pair buckets added at fixed updates (changes the data mix)

Language pairs are split into four buckets, and each bucket joins training at a fixed update count. The starts are
chosen so that pairs that overfit early join late [arxiv:2207.04672 §8.2.3]. The paper's final run enters the
buckets at 0, 170k, 230k and 270k of 300k updates. The earliest bucket is the largest, with 822 pairs, and it holds
673 low-resource pairs [arxiv:2207.04672 §8.2.3]. A shorter ablation of 100k updates entered the buckets at 0, 60k
and 80k. It gave small gains for vanilla MoE on low-resource directions, and it hurt the MoE variant that uses
expert output masking [arxiv:2207.04672 §6.2.3]. Card: [No Language Left Behind: Scaling Human-Centered Machine Translation](/papers/2207.04672).

### Sample and task sequencing in reinforcement learning (survey)

The survey orders methods by how far they change the task, from reordering experience of the final task to
defining new intermediate tasks [arxiv:2003.04960 §4.2]. At the narrow end, sample sequencing reorders experience
from the target task without changing the MDP, with prioritised and hindsight replay as examples
[arxiv:2003.04960 §4.2.1]. Co-learning lets a curriculum emerge from agents that interact or self-play in one
environment [arxiv:2003.04960 §4.2.2]. Reward and initial or terminal state changes make intermediate tasks differ
only in those properties [arxiv:2003.04960 §4.2.3]. Other methods allow any difference from the target, and
human-in-the-loop methods have people sequence the tasks by hand [arxiv:2003.04960 §4.2.4] [arxiv:2003.04960 §4.2.5].
Intermediate tasks can also be generated automatically from a parameterised domain [arxiv:2003.04960 §4.1].

Transferred knowledge is classed as low level (policies, value functions, task models, training instances) or
high level (partial policies, options, skills, shaping rewards) [arxiv:2003.04960 §4.3]. Its tables place each
listed paper on the seven dimensions, and every sample-sequencing row uses only the target task, with no transfer
[arxiv:2003.04960 tab. 2]. Card: [Curriculum Learning for Reinforcement Learning Domains: A Framework and Survey](/papers/2003.04960).

### Stage-wise task difficulty in reinforcement learning (not a data order)

A humanoid controller that follows masked motion directives is trained in three stages [arxiv:2408.07295 §4.4].
Stage 1 uses locomotion directives only, with single torso pushes. Stage 2 adds continuous pushes and grows the
episodes to 800 policy steps (16 s). Stage 3 uses whole-body directives, which may be fully specified or
partially masked [arxiv:2408.07295 §4.4]. The stages change episode length, disturbances and the mix of task
types. They do not reorder a fixed dataset.

Against a variant trained without the stages, the authors report lower simulated mimicry failure (4.875% against
0.450%) and lower walking root drift at 1 m/s (1.346 m against 0.413 m) [arxiv:2408.07295 tab. 5]. The same table
shows the opposite for standing under a push: root drift of 0.189 m against 0.256 m at 50 N, and 0.234 m against
0.311 m at 100 N [arxiv:2408.07295 tab. 5]. The comparison is not compute-matched. The variant without the stages
trained until its loss plateaued, for longer than the full model [arxiv:2408.07295 §5.5]. Card:
[Learning Multi-Modal Whole-Body Control for Real-World Humanoid Robots](/papers/2408.07295).

### Obstacle density ramp in RL (not a data order)

A quadruped locomotion policy is trained with PPO in simulation. The density of obstacles rises linearly from
sparse to dense over training, and each iteration samples a curriculum density [arxiv:2508.11849 §4.4]
[arxiv:2508.11849 alg. 1]. The paper does not state the start and end densities or the length of the ramp
[arxiv:2508.11849 tab. 3]. The ramp runs alongside per-episode physics randomisation, which is a separate
technique [arxiv:2508.11849 tab. 3]. The card's results compare architectures and token counts, not the ramp on and
off [arxiv:2508.11849 tab. 4] [arxiv:2508.11849 tab. 7], so the ramp's own effect is not isolated in the paper.
Card: [LocoMamba: Vision-Driven Locomotion via End-to-End Deep Reinforcement Learning with Mamba](/papers/2508.11849).

### Attack-strength threshold in adversarial training (the paper calls it a curriculum)

In PGD adversarial training the inner attack usually runs for a fixed number of steps on every batch. This
variant stops each example's attack once a stationarity score, FOSC, falls to or below a threshold. The
threshold falls linearly from c_max to zero over the first T′ epochs, so early epochs train on attacks that are
solved less fully and later epochs on attacks that are solved more fully [arxiv:2112.08304 §4.1]
[arxiv:2112.08304 alg. 1]. The paper treats FOSC as a difficulty measure, so it calls the method a curriculum. It
contrasts this with a curriculum that uses the PGD step count as difficulty, and argues that more steps can
overshoot [arxiv:2112.08304 §4.3]. The clean inputs stay the same, but the adversarial inputs change every epoch,
so this is a schedule on difficulty rather than a reordering of examples. In an earlier observation, replacing the
first 20 epochs of PGD with one-step FGSM gave higher final robustness [arxiv:2112.08304 §3.2] [fig. 3].

The authors report, for CIFAR-10 with an 8-layer CNN, 48.06% under PGD-10 against 44.39% for standard PGD
training and 45.54% for the step-count curriculum [arxiv:2112.08304 tab. 1]. The same run is also higher on
CIFAR-10 under FGSM, PGD-20 and C&W∞ than both baselines [arxiv:2112.08304 tab. 1]. On MNIST it has the highest
PGD-10, PGD-20 and C&W∞ values, while the step-count curriculum is higher on FGSM and on clean accuracy
[arxiv:2112.08304 tab. 1]. Under black-box attacks on CIFAR-10, the paper reports 72.15% against 65.80% for
standard training under PGD-10 [arxiv:2112.08304 tab. 2]. These are author-reported numbers. The step-count
curriculum's CIFAR-10 C&W∞ value (35.77%) is below standard training's (36.33%) [arxiv:2112.08304 tab. 1].

The convergence bound uses constants (smoothness, strong concavity, gradient variance and the inner-solve error)
that the experiments do not estimate, and the paper says Danskin's theorem does not strictly hold for ReLU networks
[arxiv:2112.08304 §4.2]. The bound therefore does not predict the reported gains. The paper suggests estimating
c_max from FGSM examples, but uses 0.5 in every run [arxiv:2112.08304 §4.1] [arxiv:2112.08304 §5.1]. Card:
[On the Convergence and Robustness of Adversarial Training](/papers/2112.08304).

### Scheduled input source in recurrent video prediction (not a data order)

During encoder steps, reverse scheduled sampling feeds the cell the true frame with a probability ε, and otherwise
the model's own previous prediction. ε rises from a start value to an end value over training [arxiv:2103.09504 §4.4]
[eq. 9]. The paper's best run uses an exponential rise from 0.5 to 1.0 [arxiv:2103.09504 tab. 6]. The paper
does not state the step counts of the schedule [arxiv:2103.09504 §4.4]. The schedule changes the inputs, not the
order of training examples, so it is a schedule in the broader sense only.

In one ablation without the decoupling loss, the exponential schedule lowers fixed Moving MNIST MSE from 57.3 to
50.6 against the model alone [arxiv:2103.09504 tab. 6]. The same schedule on a convolutional LSTM lowers MSE from
103.3 to 64.2 [arxiv:2103.09504 tab. 5]. Both results are one dataset family, and the authors report them
themselves. Card: [PredRNN: A Recurrent Neural Network for Spatiotemporal Predictive Learning](/papers/2103.09504).

### Per-class pseudo-label threshold (not a data order)

In semi-supervised image classification, the model labels unlabeled images itself and keeps only the labels it is
confident about. A single cut-off is usually used for all classes. FlexMatch gives each class its own cut-off and
moves it with how many unlabeled samples are currently assigned to that class. Classes the model has not learned
well get a lower cut-off and admit more pseudo labels. The authors call this Curriculum Pseudo Labeling
[arxiv:2110.08263 §3.1] [arxiv:2110.08263 eq. 12]. It keeps one stored class mark per unlabeled sample and adds no
parameters or extra forward passes [arxiv:2110.08263 §3.1].

The authors report lower error than FixMatch in 8 of 11 cells of their main table, with large gains on
CIFAR-100 with 400 labels (39.94% against 46.42%) and STL-10 with 40 labels (29.15% against 35.97%)
[arxiv:2110.08263 tab. 1]. It is not uniformly better.
FixMatch is equal or better on CIFAR-10 with 250 labels and on SVHN [arxiv:2110.08263 tab. 1]
[arxiv:2110.08263 §4.1]. These are author-reported results. Card:
[FlexMatch: Boosting Semi-Supervised Learning with Curriculum Pseudo Labeling](/papers/2110.08263).

### Loss-weight schedule (not a data order)

The closed-loop term of the training loss is multiplied by a weight that starts at zero and rises in steps. The
model first learns one-step prediction from real data, then learns to predict from its own outputs
[arxiv:1809.03036 §3.4] [eq. 6]. The paper does not call this curriculum learning. It changes the loss, not the
examples or their order, so it is a schedule in the broader sense only. The paper gives no step values for the
weight [arxiv:1809.03036 §3.4]. The card's source gives no number that isolates the schedule's effect; the
ablations are referred to supplementary material [arxiv:1809.03036 §5]. Card:
[A Neural Temporal Model for Human Motion Prediction](/papers/1809.03036).

### Calibrated per-entry weight, recomputed each epoch (not a data order)

Each training image in partial multi-label learning carries candidate tags, and some are wrong. Each candidate
positive is weighted by a calibrator, which maps a histogram of the model's confidences to the probability that
the candidate is correct [arxiv:2610.17635 §4.1] [eq. 7]. Each epoch, the confidences on the training set are
collected and the weights are recomputed [arxiv:2610.17635 alg. 1]. Negative entries keep weight 1 in the
algorithm [arxiv:2610.17635 alg. 1].

The calibrator is pre-trained on VOC 2007 and then used unchanged on other datasets and noise rates
[arxiv:2610.17635 §4.1] [arxiv:2610.17635 §5.2]. The authors report the best mAP in 15 of 16 noise settings against
nine baselines, with LL-R ahead at VOC with 20% noise [arxiv:2610.17635 tab. 1] [arxiv:2610.17635 §5.2]. On COCO at 80% noise, at epoch 1, the
pre-trained calibrator reaches 61.48 mAP. Fixed hard, soft and identity weightings reach 59.79, 60.08 and 59.93,
and weights from true labels reach 66.42 [arxiv:2610.17635 tab. 2]. The last row uses ground truth and is an
upper bound, not a method. The paper does not state how many runs stand behind each cell, and gives no variance
[arxiv:2610.17635 §5.1]. Card:
[Calibrated Disambiguation for Partial Multi-label Learning](/papers/2610.17635).

### Staged pre-training, alignment and adaptation (not a data order)

A vision-language-action policy for robots is trained in three runs, each on a different data set and with
different parts trainable [arxiv:2502.05855 §3.2]. Stage 1 trains only the action expert on cross-embodiment
data, about 100 hours over 91 tasks and four robot bodies [arxiv:2502.05855 app. C.4]. Stage 2 keeps one
embodiment and trains the language model, the projection and the expert together. Stage 3 fine-tunes on
demonstrations for one task, and the authors skip it for simpler tasks [arxiv:2502.05855 §3.2]. The card does
not describe these stages as a curriculum. The card lists the paper under this concept, and the stages move from
broad to specific data, so this is a schedule in the broader sense only.

In the authors' stage ablation, runs without stage 1 or without stage 2 score 0.0 on both shirt and laundry
folding. Without stage 3, laundry folding scores 0.0 against 0.4 with it, and shirt folding is 0.92 in both
[arxiv:2502.05855 §4.4]. The card notes that the paper's explanation for the stage-1 result (that stage 1
warms up a large expert) is a hypothesis the paper does not test [arxiv:2502.05855 §4.4]. These are author-reported results on the
authors' own robots, with 10 trials per task and no variance given [arxiv:2502.05855 §4.1] [arxiv:2502.05855 §4.3].
Card: [DexVLA: Vision-Language Model with Plug-In Diffusion Expert for General Robot Control](/papers/2502.05855).

### Scale-stage growth in NeRF (adds data and a network block per stage; not a data order)

A neural radiance field for city scenes seen from satellite to street level is trained in four stages
[arxiv:2112.05504 §3.3]. Stage 1 uses only the most distant image scale and a four-layer base block. Each later
stage adds the next closer scale to the training rays and appends one two-layer residual block with its own output
head [arxiv:2112.05504 §3.3]. Head L is trained on every scale up to L, so each head gives one level of detail
[arxiv:2112.05504 eq. 3]. The stages change the training set and the network together. The set grows rather than
being reordered, so this is a schedule in the broader sense only. The paper's reason is that training all scales at
once biases a plain NeRF toward low frequencies [arxiv:2112.05504 §3.2].

The paper gives 100,000 iterations per stage [arxiv:2112.05504 §4]. Its main comparison uses a row labelled "same
iter as baselines", which the paper says uses the baselines' budget but does not give as a number. Its
"until convergence" row trains longer [arxiv:2112.05504 tab. 2]. On average PSNR, the same-iteration row gives 23.48
on 56 Leonard and 23.61 on Transamerica, against 22.31 and 22.83 for the strongest Mip-NeRF baseline
[arxiv:2112.05504 tab. 2]. In the ablation, dropping the data-feeding scheme gives 22.34 on 56 Leonard and 23.02 on
Transamerica, and dropping model growing gives 22.14 and 23.06 [arxiv:2112.05504 tab. 3]. On the averages, the
no-feeding variant is only just above the strongest baseline on both scenes, which contradicts the text's wording
[arxiv:2112.05504 §4.2]. The tables cover two scenes, and the authors report them themselves. Card:
[BungeeNeRF: Progressive Neural Radiance Field for Extreme Multi-scale Scene Rendering](/papers/2112.05504).

## When to use which

| Situation | Use | Why |
|---|---|---|
| No extra model or parser, and a tight budget | A simple complexity sort, such as length | Cheap to build. One submission's complexity-based curricula showed no consistent gain over no curriculum or reversed order [arxiv:2504.08165 app. F] |
| A teacher model is affordable and the extra pass is acceptable | Surprisal-based ranking, followed by random-order training | The LSTM-teacher run beat its baseline only after a random-order phase. The n-gram run gained on GLUE but lost on MSGS [arxiv:2504.08165 app. F] |
| Linguistic annotation is available | Syntactic-complexity ordering | Results conflict: one team beat random order on the 10M-word track [arxiv:2504.08165 app. F], while another found all its curricula below a random baseline [arxiv:2504.08165 app. F] |
| An architecture, objective or preprocessing change is an option | Try these first | The paper reports these as the most consistent gains [arxiv:2504.08165 §9] |
| Sequence length is a tunable setting | Tune the context length | One team's ablations found sequence length mattered [arxiv:2504.08165 app. F]. Another found a context of 32 beat the baseline's 128 [arxiv:2504.08165 app. F] |
| Semi-supervised image classification with few labels, on a FixMatch-style pipeline | Per-class thresholds that move with learning status | Lower error than FixMatch in 8 of 11 cells of the authors' table, with the largest gains on STL-10 with 40 labels and CIFAR-100 with 400 labels. FixMatch is better on SVHN and on CIFAR-10 with 250 labels [arxiv:2110.08263 tab. 1] |
| A simulated robot controller with many skills, and budget for a second, no-stage training run | Stage-wise difficulty, from locomotion to whole-body directives | Lower mimicry failure and walking drift in the authors' ablation, but worse standing drift under push, and the comparison is not compute-matched [arxiv:2408.07295 tab. 5] [arxiv:2408.07295 §5.5] |
| Locomotion trained with PPO in simulation, with an obstacle or terrain density you can set per iteration | A linear ramp from sparse to dense obstacles, combined with physics randomisation | The paper uses the ramp with physics randomisation, but its tables compare architectures, not the ramp on and off, so the ramp's own effect is not isolated [arxiv:2508.11849 §4.4] [arxiv:2508.11849 tab. 4] |
| Adversarial training with PGD, and a per-example stopping rule for the inner attack is acceptable | A FOSC threshold that falls linearly to zero over the first epochs | Higher PGD-10 and PGD-20 than standard training and the step-count curriculum on CIFAR-10 with an 8-layer CNN [arxiv:2112.08304 tab. 1]. On WideResNet, PGD-20 is higher than Madry's setting but clean accuracy is lower [arxiv:2112.08304 tab. 3]. The paper sets c_max and T′ by hand per dataset [arxiv:2112.08304 §5.1] |
| Recurrent predictor trained with teacher forcing | A reverse scheduled sampling schedule that rises to 1.0 | Lower MSE in the authors' ablations on fixed Moving MNIST, across several backbones [arxiv:2103.09504 tab. 5] [arxiv:2103.09504 tab. 6]. Only one dataset family is tested |
| Noisy multi-label labels, and a calibrator can be pre-trained once | A calibrated per-entry weight, recomputed each epoch | Best mAP in 15 of 16 settings in the authors' table, but LL-R leads at VOC with 20% noise [arxiv:2610.17635 tab. 1] |
| A robot policy with a large pre-trained vision-language backbone, several robot bodies in the pretraining data, and a few demonstrations for the target task | Three runs: broad pretraining, then one embodiment, then task fine-tuning | The authors' stage ablation gives 0.0 without stage 1 or stage 2, and stage 3 lifts laundry folding from 0.0 to 0.4. Results are on the authors' own robots with 10 trials per task [arxiv:2502.05855 §4.4] [arxiv:2502.05855 §4.1] |
| Many language pairs of uneven size, one large multilingual model, and a fixed update budget | Split pairs into buckets and add the later buckets at fixed updates | On the authors' 128-expert ablation, the average is 47.6 with no curriculum, 48.4 with naive bucketing and 48.5 with four phases [arxiv:2207.04672 tab. 29]. The gain over no curriculum is under one point, and the shorter ablation found the curriculum hurt the MoE variant with output masking [arxiv:2207.04672 §6.2.3] |
| Scenes seen at very different camera distances, with several levels of detail wanted from one NeRF, and a budget for several training stages | Add the next closer scale and one residual block with its own head at each stage | Average PSNR above the Mip-NeRF baselines at the same-iteration row in the authors' table, though that budget is not given as a number. The no-feeding ablation is close to the strongest baseline on averages [arxiv:2112.05504 tab. 2] [arxiv:2112.05504 tab. 3] |
| Reinforcement learning, and only the target task's experience may be reordered | Sample sequencing, such as prioritised or hindsight replay | The survey keeps the target MDP and lists these papers as single-task curricula with no transfer [arxiv:2003.04960 §4.2.1] [arxiv:2003.04960 tab. 2] |
| Reinforcement learning, and intermediate tasks may differ from the target in any way | Pick by the transfer method and ordering you can afford | The survey's classification is mixed across sequence and graph curricula, value and policy transfer, and static and adaptive ordering [arxiv:2003.04960 tab. 2]. It gives no ranking |

The data-order evidence rests on one paper, so the first five rows do not rank methods on general grounds. Any
choice should be checked against a random-order run on your own data. The per-class threshold, attack-strength threshold, stage-wise,
obstacle-ramp, scale-stage, input-schedule, calibrated-weight, staged-pipeline and bucket rows each rest on one paper's own experiments, so
check them against a fixed-schedule baseline on your data. The loss-weight schedule has no comparison in the card to base a choice on,
so it is not in this table. The last two rows are classifications of published work, not measured comparisons, so
they say where a method sits and do not say which one wins.

## Implementation sketch

Language-agnostic pseudocode of the common core. Variants differ in the score and in how many phases they use.

```pseudocode
# Curriculum over a fixed corpus  [§7.4] [app. F]
# Input:  corpus (list of examples), score(example) -> complexity
# Output: one training order over the same examples

def curriculum_order(corpus, score, phase_two_random=True):
    # Rank every example by one complexity measure (length, frequency,
    # surprisal, parse depth) [§7.4]
    ranked = sorted(corpus, key=score)
    # Easy-to-hard first phase
    if not phase_two_random:
        return ranked
    # Second phase: a random permutation of the same examples [app. F]
    return ranked + random_permutation(corpus)
```

Notes: the sort key is the only part that changes across variants. The two-phase tail is one reported pattern,
not the only one [arxiv:2504.08165 app. F]. Surprisal ranking needs a scoring model, an LSTM ensemble or an
n-gram model, run over the whole corpus before training starts [arxiv:2504.08165 app. F]. The loss-weight
schedule is a different operation (a per-step weight on the loss term) and is not covered by this sketch
[arxiv:1809.03036 §3.4]. The per-class threshold is also a different operation (a class-dependent mask on
pseudo labels) [arxiv:2110.08263 §3.1]. The bucket schedule adds whole groups of examples at fixed update counts,
which this sketch does not cover [arxiv:2207.04672 §8.2.3]. Scale-stage NeRF training also grows the network at each
stage, adding a block and an output head, so it does not fit this sketch either [arxiv:2112.05504 §3.3]. The stage-wise RL curriculum changes the episode length, the disturbances
and the directive sampler, not a list of examples [arxiv:2408.07295 §4.4]. Reverse scheduled sampling is a per-step
choice of input, made with probability ε [arxiv:2103.09504 eq. 9]. The calibrated weight is a per-entry loss weight
set from confidence bins, recomputed every epoch [arxiv:2610.17635 alg. 1]. The DexVLA stages are separate training
runs over different data and trainable parts, so they are not a sort of one corpus [arxiv:2502.05855 §3.2].
The obstacle ramp changes the simulated terrain, not a list of examples [arxiv:2508.11849 §4.4]. The attack-strength
threshold stops each example's inner attack on a per-example score inside the adversarial loop; the example set is
unchanged [arxiv:2112.08304 alg. 1].

## Contradictions

- **Majority negative vs. positive summaries.** The paper says most curriculum attempts did not improve
  consistently [arxiv:2504.08165 §7.4], but also says one curriculum method gave significant gains
  [arxiv:2504.08165 §9] and that curricula can beat a naive ordering on GLUE [arxiv:2504.08165 app. D].
- **Linguistic ordering.** One team reported that six linguistic complexity metrics beat a random order on the
  10M-word track [arxiv:2504.08165 app. F]. Another found dependency-based ranking best among its curricula, yet
  all of them underperformed a random baseline [arxiv:2504.08165 app. F].
- **Surprisal ordering.** The LSTM-teacher version did not beat a random baseline on its own, but did when
  followed by random-order training [arxiv:2504.08165 app. F]. The n-gram version improved GLUE but not MSGS
  [arxiv:2504.08165 app. F].
- **Curriculum with output masking.** The final NLLB model combines the four-phase bucket schedule with expert
  output masking [arxiv:2207.04672 §8.2.4]. The authors' shorter ablation found the curriculum helped vanilla MoE on
  low-resource directions but hurt MoE with output masking [arxiv:2207.04672 §6.2.3]. The two runs differ in length
  (100k against 300k updates), so the paper does not settle the question.
- **MSGS.** The paper says principled curriculum approaches can improve the MSGS score [arxiv:2504.08165 app. D],
  while its main summary says most curricula did not help [arxiv:2504.08165 §7.4].
- **Scale-stage NeRF, text against table.** The paper says that without progressive data feeding the results are
  "even inferior than baselines" [arxiv:2112.05504 §4.2]. Its averages in tab. 3 show the variant slightly above
  the strongest baseline on both scenes [arxiv:2112.05504 tab. 2] [arxiv:2112.05504 tab. 3].
- **Stage-wise RL, by setting.** The same ablation favours the staged curriculum on walking and mimicry, and the
  no-stage variant on standing under push [arxiv:2408.07295 tab. 5]. The authors' text calls the standing
  performance on par with the baseline, while the baseline comparison table shows a larger drift for the controller
  under a 50 N push [arxiv:2408.07295 §5.3] [arxiv:2408.07295 tab. 2].

## Papers

- [Findings of the BabyLM Challenge: Sample-Efficient Pretraining on Developmentally Plausible Corpora](/papers/2504.08165):
  the first BabyLM shared task, with 31 submissions and a meta-analysis that reports most curriculum attempts
  did not improve consistently.
- [FlexMatch: Boosting Semi-Supervised Learning with Curriculum Pseudo Labeling](/papers/2110.08263): applies a
  per-class, learning-status-dependent threshold to FixMatch-style pseudo labels. The name refers to curriculum
  learning, but the schedule acts on the pseudo-label threshold, not on the order of examples.
- [A Neural Temporal Model for Human Motion Prediction](/papers/1809.03036): listed under this concept for its
  closed-loop loss-weight schedule. The paper does not describe that schedule as curriculum learning.
- [Learning Multi-Modal Whole-Body Control for Real-World Humanoid Robots](/papers/2408.07295): a three-stage
  difficulty curriculum for one humanoid controller, with an ablation against a variant trained without it. The
  comparison is author-reported and not compute-matched.
- [BungeeNeRF: Progressive Neural Radiance Field for Extreme Multi-scale Scene Rendering](/papers/2112.05504): a NeRF
  that adds one closer image scale and one residual block per stage, with an output head per level of detail. Its
  ablations are author-reported and cover two city scenes.
- [PredRNN: A Recurrent Neural Network for Spatiotemporal Predictive Learning](/papers/2103.09504): uses reverse
  scheduled sampling, a schedule on the input source rather than on the examples. Listed under this concept by the
  card.
- [No Language Left Behind: Scaling Human-Centered Machine Translation](/papers/2207.04672): a four-phase schedule
  that adds language-pair buckets at fixed updates for a 54.5B-parameter MoE translation model. Its own ablations
  are author-reported and use the dev set.
- [Calibrated Disambiguation for Partial Multi-label Learning](/papers/2610.17635): a calibrated per-entry weight
  recomputed each epoch, which the paper calls a curriculum loss. Its weights are not a data order.
- [Curriculum Learning for Reinforcement Learning Domains: A Framework and Survey](/papers/2003.04960): a survey that
  defines curriculum learning for RL as task generation, sequencing and transfer, and classifies published methods
  on seven dimensions. It has no experiments of its own.
- [DexVLA: Vision-Language Model with Plug-In Diffusion Expert for General Robot Control](/papers/2502.05855): a
  three-stage training pipeline (cross-embodiment pretraining, embodiment alignment, task adaptation) for a robot
  policy. Listed under this concept by the card. Its results are author-reported on the authors' own robots.
- [LocoMamba: Vision-Driven Locomotion via End-to-End Deep Reinforcement Learning with Mamba](/papers/2508.11849): a
  PPO locomotion policy in PyBullet that ramps obstacle density over training. The paper's comparisons are
  architecture comparisons, and the ramp is not isolated. Results are author-reported.
- [On the Convergence and Robustness of Adversarial Training](/papers/2112.08304): a FOSC threshold that falls
  linearly over training, which stops each PGD example once its inner solve is good enough. The paper calls this a
  curriculum, but the threshold sets attack strength, not example order. Results are author-reported.