acceptodds
Under review as a conference paper at ICLR 2027

Level Up: Defining and Exploiting Transitional Problems for Curriculum Learning

Abstract

Curriculum learning—ordering training examples in a sequence to aid machine learning—takes inspiration from human learning, but has not gained widespread acceptance. Static strategies for scoring item difficulty rely on indirect proxy scores of varying quality and produce curricula that are not specific to the learner at hand. Dynamic approaches base difficulty estimates on gradient information, requiring considerable extra computation during training. We introduce a novel method for measuring the difficulty of individual problem instances that is calibrated to a series of models of increasing competence, and identify *transitional problems* that are consistently easier as model ability increases. Applying this method to diverse model series constructed from sets of models that are readily available on many tasks, we find that training on a curriculum that *levels up* from easier to harder transitional problems most efficiently improves a model to the next tier of competence. These problems induce a natural progression from easier to harder items, which outperforms other training strategies. By measuring difficulty directly relative to model competence, our method yields interpretable problems, learner-specific curricula, and a principled basis for step-by-step improvement.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Curriculum Learning, Difficulty, Hardness, Level, Transitional Problem, Static Curriculum, Model Series

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_2tecOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Summary

The paper presents an investigation of a new curriculum learning approach. The approach entails the "bucketing" of data points into buckets of increasing difficulty. Different to other methods, they define this difficulty by the "capability index" of the least capable model that can correctly predict them. They have a reasonable definition of this capability index congruent with current understanding (it basically entails looking at model size or number of few shot examples).

Recommendation

It's not stellar and there is on experimental limitation that might cause me to revise my score downwards or upwards (see below), but this is an interesting investigation on a very interesting topic with a fresh take and a very clear story. The language is clear and the paper is easy to follow. The authors have left themselves very little space to interpret their findings (discussion is only 1 paragraph long), which I think is a major downside of this paper (see below).

At this stage, I will give this paper a "weak accept" because I have some misgivings about some aspects, but all in all I think it deserves publication and is a nice addition to the literature on curriculum learning

(Even if the paper is not accepted now and even if the issue I identify with evaluation is indeed true, I still recommend the authors revise and try to publish elsewhere because the idea is truly interesting and I think the community needs to see the results)

Reviewer self-evaluation

I have been working on curriculum learning on-and-off for the last 5-6 years, primarily on domain-specific aspects of it relevant to my own field. My field is speech (primarily audio, less so text). I am thus very familiar with NLP methodologies and curriculum learning, though I haven't done much work on the intersection of the two.

Critical strengths

New definition of difficulty

The authors give a new and very interesting definition of instance difficulty (the score that is used to rank examples during curriculum learning)

They define it based on model competence, evaluating three definitions of competence:

  1. Model size
  2. Checkpointing (i.e. training for longer/shorter)
  3. Length of the context window

They then consistently show that

  1. Easy-to-hard curricula using their metrics outperform IID training and relevant curriculum learning baselines
  2. Hard-to-easy curricula underperform IID (though not all other easy-to-hard baselines in all cases)

They do so on two domains (math and chess)

This is a very interesting finding from both a practical and a theoretical perspective:

  • Practical perspective: It helps train smaller LMs better (also faster? but this doesn't seem to have been checked..)
  • Theoretical perspective: I think this has interesting repercussions for understanding these models and problems. Alas, this is where I think that the authors erred by investing too little time in interpretation (see below)

A clear story

I said this before but I want to re-emphasize it (I guess I'm biased because I just reviewed too many papers who tried to do too many things at once) -- the authors have a clear story from beginning to end. All experiments serve that story. There are no random or irrelevant side-quests. The reader is guided into an investigation and leaves it with more information than they came in. In my case, as someone interested in curriculum learning, this also left me with more ideas for my own research and more questions that I would like to ask the authors during the conference (I probably won't be there, I hope that the authors ultimately make it).

In short, I believe this is precisely the type of paper that should be published at ICLR..

Alas, I have some misgivings which should be addressed first...

Critical weaknesses

No space for discussion

I understand that the page limitation is too limiting (everyone seems to be struggling with this), but it's there for some reason and authors are supposed to take it or leave it (I also struggle with this)

So on the one hand, the authors have to fill in a lot that deserves to be there, but on the other hand, they're left without any space to interpret their findings.

However this is precisely the thing we readers are waiting for! The authors do have plenty of interesting observations. Do they really need Figure 1? Can they not cut some parts of the introduction or related work, which is indeed thorough and relevant but doesn't leave any space for the more interesting part?

Take Figure 10 for example. The authors note that LevelUp (easy-to-hard) starts with a much larger gradient than IID and gradually decays towards it, while LevelDown (hard-to-easy) starts close and deviates later.

The authors then go on to note that "[...], indicating that an effective curriculum exposes the learner to a qualitatively different optimization problem."

Different how??!?!?!?!? This is precisely what the reader wants to find out. Why do "easy" examples have a higher gradient norm? And why do most learning rate recipes for transformers "warm up" the learning rate, thus starting from softer updates and gradually ramping up, as LevelDown appears to be doing. Is there any connections there?

There is so much that can be unpacked from these observations and it is a pity that the authors leave themselves with just enough space to say that "we tried it and it worked". I strongly suggest them to find a way to get more interpretation into their work (a very hard task, I know!)

Figure 6 vs Human Curricula

Figure 6 is the only attempt to give some depth to the interpretation.

There seems to be a monotonic relationship between "Average Question Length" and "Levels" (although the figure is missing error bars!)

So, on first interpretation, those two curricula should be interchangeable if the same pacing rule is applied. Yet LevelUp consistently outperforms the human curricula, making me wonder if the pacing was exactly the same.

Specifically, given that you essentially create hard "buckets" using the transitional problem definition, which gives you a discrete difficulty value, how do you handle curricula with "softer" values, like the question length? Because if you apply a split based on Figure 6 (leftmost), then you should get identical data buckets as for LevelUp, right? Then you should also get the same performance..Unless the error bars are too high and then the ordering is indeed different..

Questions for authors

Like I said in the introduction, I would appreciate an answer to my question regarding Figure 6 and Human Curricula. This could cause me to change my score both upwards and downwards

Beyond that, I don't have a concrete question, but kindly ask the authors to consider the broader repercussions of their findings, and perhaps to state them here (and/or revise their discussion accordingly). What are we to make of this?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 2tec

We thank the reviewer for their thoughtful and detailed review on our work. We are grateful to the reviewer for accurately capturing the spirit of our work, and for their positive comments and acceptance recommendation. We agree that our results merit additional discussion and present an extended analysis of our results below, with an emphasis on the reviewer’s primary concern: our interpretability figures. We hope that our direct and detailed response will positively influence the reviewer’s recommendation of our work.

Extended Discussion

Overall, our objective in this work is to adapt multiple successful ideas from human curricula — the tiered education system and the Zone of Proximal Development (ZPD) — into a machine learning system. As we outline in section 1, human curricula are successful in part because they are calibrated to the individual needs of weaker and stronger learners, and because they emphasize what is just out of reach. By defining the transitional training problems that characterize this specific notion of difficulty for a learner, we are able to improve upon existing curriculum strategies even with a simple, linearly structured training regime.

  • This approach directly connects to our result in figure 10. The qualitative difference is as follows: the gradient norm is a weak proxy for the learning signal in a model. An objectively easy problem that the model can already solve will lead to minimal need for change (small gradient norm, high performance), and a very hard problem may make any gradient step lead to little improvement, again resulting in a small gradient norm everywhere.
  • In our experiment, the learner (Maia-2 level 1000) fails every transitional problem by construction. Level-1 problems target a narrow gap shared across problems (1000 to 1100), so per-example gradients agree and add up in the batch. High-level problems need skills far from the learner, so their gradients point in different directions and partly cancel, giving a norm close to IID. This explains both the LevelUp and LevelDown curves.
  • Finally, aside from being computationally expensive, directly selecting based on the gradient norm greedily selects samples that may be easy to learn but not beneficial for generalization. Thus, our method also calibrates for problems that stronger learners are actually able to solve, leading to a successful and smoothly improving LevelUp curriculum.

Analysis of Figure 6 vs. human curricula

  • The reviewer’s intuition for the differences between the length-ordering and LevelUp is exactly right. The boxplots in Section R4 of our response experiments (https://anonymous.4open.science/r/level-up-rebuttal-363A/) show that there is a high variance in our interpretability results (length, steps, etc.), even though the means are correlated with the transitional level.
  • The reason for this variability is that the difficulty of a problem may strongly correlate with length, but this does not mean that all long problems are qualitatively similar. For example, consider the problems: 'multiply 2 by 4' and 'what is 236 / 41?'. Though they are of a similar length, one is significantly harder than the other. Though the problems at each transitional ‘bucket’ vary monotonically in length on average, their composition can vary quite sharply. We show that this is indeed the case for GSM8k problems, with concrete length-matched examples from each level (Section R4).
  • Our length-based curriculum method performs a stratified sampling from lengths 1 to max_len, rather than just picking the shortest few. Thus, both methods use linear pacing. Since a practitioner would not have access to the min and max length of the buckets that form transitional problems without defining and measuring a model series, we assert that this is a fair comparison to a length-based curriculum with the same training budget.

On the broader repercussions of our work

We present a novel view on the connection between curricula designed by human teachers and those for ML models. Even within the same syllabus, teachers/tutors often calibrate their exact curricula to what a weak or strong student may learn. Our method is in a similar spirit, as it modulates the data diet to a model based on what it performs poorly on, but is still achievable by other models. This is an interpretation of the Zone of Proximal Development (ZPD) concept in psychology that we describe in Appendix A. We will add this and the above discussion to the final version of our paper, and we thank the reviewer for sparking this discussion which adds value to our findings.

Reviewer_2tecOpenReview ↗

Not convinced, keeping my score

Thanks for taking the time to respond, but I'm not (yet?) convinced to change my score.

  1. I only see a very loose connection with ZPD. In fact, it's so loose that it can be applied to all curriculum learning methods, thus rendering it almost vacuous. You are using the difficulty scores of progressively more capable models to identify easy-to-hard rankings of the training data. This does not say anything about whether harder samples are outside of the ZPD for the current model. (It also doesn't say anything about how this should be measured.) For instance, you do not evaluate whether the model agrees with the ranking of the other models or so.

  2. Figure 6 vs human curricula: Okay, the boxplots appear to have a wider variability than I expected. Isn't that the main takeaway? So doesn't that invalidate what you were trying to say there?

Anyways, these answers don't convince me to increase my score further, but the also don't cause me any additional concerns, so I'll keep it the same.

AuthorsOpenReview ↗

Clarification on ZPD connection + Figure 6

Thank you for your follow-up questions. We greatly appreciate your continued engagement and the opportunity to discuss our work in detail. Below we further clarify our stance on these two aspects of our work, and hope you will consider raising your score.

1. Connection to the ZPD

  • We hypothesize that transitional problems represent the ZPD because they are the problems that distinguish one competence level from another — they are the problems that are just out of reach of a model at one level but crucially, solved by the next levels of models.
  • In contrast, static domain-specific curricula do not explicitly order difficulty in a way that's specific to the learner. Dynamic curricula capture some notion of the ZPD, but aside from being expensive to run, they are short sighted, and thus what is out of reach may not actually represent the ideal direction in which development (i.e., improvement towards a stronger model) should occur.
  • Our problem ranking relies on the full series, and changing the order of the series will change the ranking. Since we order by performance on the target task, transitional problems represent a progression in competence relative to the learner (weakest model), following the idea of scaffolding/the ZPD.

2. Two key takeaways from Figure 6

  • Our main goal in showing Figure 6 is that our procedure for selecting transitional problems on a task, which is derived entirely from the performance of weak and strong models on that task, is still weakly aligned with some measures of human difficulty. This is neither a perfect recreation (which would be shown by a very narrow boxplot) nor is it a complete lack of alignment (which would be shown by an uncorrelated mean).
  • Thus, the trends in Figure 6 and the boxplots can be interpreted in two ways: as a 'sanity check' that higher levels of transitional problems do align with our intuition of what is hard, but also as a way to show that they can't be reconstructed from a few such domain-specific measures alone. The boxplots demonstrate both of these points, and we thank the reviewer for highlighting this result.
Reviewer_FkN8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The manuscript introduces transitional problems identified from models with increasing task ability and uses them to build an easy-to-hard curriculum that can help train models more effectively.

Critical strengths

Strengths

  1. The study addresses an important question in curriculum learning that how to define the difficulty of a problem for a model. The introduction explains why human-designed difficulty measures may not fit model learning and why methods that update difficulty during training can be costly.

  2. The definition of transitional problems is clear. A problem belongs to a level when weaker models fail to solve it and all models at or above that level succeed. It is easy for readers to understand.

  3. The main outcome is encouraging. In the Qwen2.5-0.5B model-family experiment, LevelUp reaches 21.40% test accuracy on transitional problems, compared with 16.67% for random training.

  4. The study evaluates the method in multiple experiments. The research study transfer to processed Orca problems and conduct chess experiments, which provide additional evidence that transitional problems can be useful in broader situations.

Critical weaknesses

Weakness:

  1. The study assigns equal attention to each level, but the number of problems differs across levels. Problems in levels with fewer examples may therefore be seen more often, which may affect the comparison with the uniformly sampled RANDOM baseline.

  2. The study defines a binary correct-or-wrong label for each model and problem, but math performance is evaluated using 8 answers per problem. It is confusing how these answers are transformed to binary labels.

  3. The study argue that the problem levels can only be computed once and then be reused. However, does this really reduce the overall computational cost? Suggest report the cost of computing the levels and comparing the total cost with that of the baselines.

  4. In one additional experiment, the result for LevelUp is better than that for random training, but almost the same as that for IID. The authors should clarify the boundary conditions under which LevelUp is useful.

  5. In the math experiments, the models are trained for only 8-15 steps. Although the study compares some training budgets, longer training is needed in more settings. This would demonstrate whether LevelUp only helps the models learn faster at the beginning or also improves their final performance.

Questions for authors

Please find my questions from weaknesses section.

2 comments
AuthorsOpenReview ↗

Response to reviewer FkN8

We thank the reviewer for their thorough assessment of our work. The reviewer accurately captures the essential details of our work and raises important questions that help to clarify our contribution. We address each of the reviewer’s questions point-wise below, and we hope that the reviewer will consider recommending acceptance as these clarifications strengthen the robustness of our work.

1. Variation in problems across levels

  • In most of the settings we explore, each level contains a sufficient number of unique problems to fill out a training block. The number of steps in the GSM8k experiments ensures no repetition, and we have a large number of problems in the chess setting.
  • The IID baseline on the full dataset does not assume any additional structure, but we do provide a setting in which the problems may be sampled multiple times based on the size of a level: IID on transitional problems. Since we randomly sample across levels, the sampling is non-uniform w.r.t. the full dataset.

2. Binary correctness in the sampling-based setting

  • First, we note that only the math setting requires sampling, as in the chess setting, we always pick Maia-2’s most probable next move.
  • In our GSM8k experiments, we use a single sample to determine the transitional problems. While this leads to sampling variability, we show in Section R2 of our experimental rebuttal (https://anonymous.4open.science/r/level-up-rebuttal-363A/) that transitional problems are still aligned with pass rates for each level of model. Further, curricula based on sampling pass rate do not outperform LevelUp in general.
  • Additionally, we show an example of binarizing a multi-sample evaluation in our new GRPO experiments, which are conducted with the advanced Qwen3.5 models on the more challenging DeepScaleR math dataset (Section R3). Finding transitional problems by binarizing via pass rate >= 0.5 still leads to a successful LevelUp curriculum in this complex setting.

3. Computational cost and reuse

  • We report the cost of collecting transitional problems in the math setting in Table 5. While this cost is greater than that of one training run, we note that most practical scenarios involve multiple training runs (e.g., for hyperparameter search).
  • We also show that transitional problems can transfer from a small dataset to a large one (see the experiments on the Orca dataset). In the Orca setting, a single training and evaluation pipeline takes over 30 minutes, which is higher than the cost of collecting and transferring neo-transitional problems. This further amortizes the cost of transitional problems as the dataset sizes and training budgets increase.
  • Since each model in the series has to be evaluated only once and the evaluations are cheap to store (one numerical dataset column per model), transitional problems can be constructed from multiple sub-series without re-evaluation.

4. Boundary conditions for LevelUp usefulness

  • We expect LevelUp to be useful when there is a distinct difference in performance between the models in a series. Models that are very similar in strength (e.g., with weak in-context adaptation) will produce less differentiated problems than those with strong differences (e.g., family of LLMs), and thus provide a smaller benefit to learning.
  • Further, as we describe in Section 2 of the paper, existing work shows that curricula are primarily beneficial for learning speed, and when data are noisy or complex. Our primary objective is to show an improvement in performance compared to other curricula, and we are therefore subject to the general limitations of curriculum learning outlined in these analyses.

5. Long training budgets in math

  • Though we only use a few steps in the math setting, our experiments use a large batch size of 64, and thus train on hundreds of examples. We choose the training budget so that we do not repeat any transitional problems during training.
  • To extend our results to a more complex math domain with larger budgets, we run each curriculum on the Qwen3.5-0.8B model with GRPO on the DeepScaleR math dataset. Even with 300–500 steps of RL (8h+ of training on our hardware), we observe that LevelUp outperforms all baselines. We note that the model is still improving in all cases, and we expect IID to eventually catch up. However, this highlights that LevelUp is beneficial even with significant computational investment. Results and details are presented in Section R3 of our experimental rebuttal.

I thank the authors for the response. I maintain my rating.

Reviewer_8zp8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a method for measuring the difficulty of individual problems relative to a series of models with increasing competence, and uses it to identify “transitional problems.” It shows that curriculum learning on these problems, ordered from easy to hard (LevelUp), is more effective than other orderings.

Critical strengths
  • The idea is easy to understand, and the paper explores several ways to define problem difficulty (model families, checkpoints, k-shot prompting, and skill-conditioned models).
  • The results on intra-model transitional problems in chess are interesting.
Critical weaknesses
  • The paper constructs transitional problems in several ways, and the relationship between the model series and the learner differs across experiments. For example, the levels are sometimes defined by checkpoints of the same model, sometimes by larger models, and sometimes by the same weights with a different skill input (Maia-2). Because of this, it is unclear where the gain comes from. Is it from the transitional problems themselves, from the monotonicity constraint, or from the specific relationship between the series and the learner? An analysis that separates these factors would make the main claim clearer.
    • The only model-based baseline is the C-score, and comparisons with other model-based curricula are missing. Some of these can be built from the same correctness matrix at almost no extra cost, for example, 1) sorting by the pass rate of a single strong model or 2) sorting by the learner’s own loss or pass rate.
    • Comparisons with recent curriculum methods [1, 2] would also be helpful.
  • The GSM8K performance seems quite low compared to the officially reported numbers for the Qwen2.5 base models.
  • The training budget for the math experiments (fewer than 15 steps) seems very small.
    • As the authors mention in the introduction, Wu et al. found that curricula often stop helping with larger training budgets, and recent work reports similar findings for LLM reasoning [3]. Figures 17 and 20 also suggest that the gap shrinks as training continues. If a model trained with LevelUp is further trained with a longer budget, does the gain still hold?
    • If not, what is the main benefit of LevelUp?
Questions for authors
  • When computing the matrix, do you use a single inference per problem? Since evaluation uses sampling (temperature 0.75), a model may answer the same problem correctly in some runs and incorrectly in others, which could cause large variance. If multiple samples are used, how are they converted into a binary label?
  • How did you train the model on the math datasets? What target and loss did you use?
1 comment
AuthorsOpenReview ↗

Response to Reviewer 8zp8

We thank the reviewer for their meticulous review, and for recommending ablations and new experiments that help to additionally contextualize our work. Below, we conduct an extensive analysis based on the reviewer’s suggestions, and respond to each of their critiques point-wise. These additions strengthen our contribution, and we respectfully request the reviewer to consider recommending acceptance to ICLR 2027.

The gain from LevelUp

The gain from LevelUp stems from all three components: the ordering of the model series, the scores of the model series on a problem, the consistency constraint that calibrates the difficulty of a problem relative to more and less competent models. We detail the contributions of each component below.

  • As we show in Section R1 of our response experiments (https://anonymous.4open.science/r/level-up-rebuttal-363A/), ablating the monotonicity constraint causes a 1–2% decrease in LevelUp performance. Thus, this constraint adds an important learning signal, but its absence does not collapse the advantage of our construction over IID.
  • The specific relation of the model series to the learner does matter (e.g., bigger competency gaps can provide more of a learning signal). However, by showing positive results across diverse series, our main goal is to establish that the LevelUp curriculum is useful to practitioners in a general setting. A strong example of this is the Maia-2 model series: every model in the series shares the same weights and only the skill input changes, yet LevelUp still wins.
  • To our knowledge, the C-score is the only existing difficulty measure that aggregates the performance of diverse models. In the chess setting, we always pick Maia-2’s most probable next move, so there is no sampling. In our GSM8k experiments where model outputs are sampled, we compare to the suggested pass-rate curricula (Section R2), and find that LevelUp performs at least as well as the pass rate curricula in every setting, and is significantly better than each in at least one of these settings.

Computing and curricula in the sampling-based setting

  • In our GSM8k experiments, we use a single sample to determine the transitional problems. While this leads to sampling variability, we show in Section R2 that transitional problems are still aligned with pass rates for each level of model. Further, the suggested pass rate curricula do not outperform LevelUp in general.
  • Additionally, we show an example of binarizing a multi-sample evaluation in our new GRPO experiments (Section R3). Finding transitional problems by binarizing via pass rate >= 0.5 still leads to a successful LevelUp curriculum in this complex setting.

Qwen2.5 on GSM8k: baseline score and training

  • The official evaluation of the Qwen2.5 models reports accuracy as pass@1 with 4-shot prompting, i.e., with 4 in-context examples. Here we report 0-shot mean@8, a more challenging setting that leads to lower numbers. In Section R0 of our response we show that our exact evaluation pipeline recovers the officially reported numbers with 4-shot ICL, with small variations due to prompting differences.
  • Training. We trained the Qwen2.5 models with SFT on the math datasets, using the gold solution from the GSM8k dataset as the target. We used the autoregressive objective (i.e., per-token teacher forcing with a cross-entropy loss) as is standard on this dataset. In our new experiments (Section R3), we train using GRPO (8 generations/problem, i.e. 48 samples/step).

On the effectiveness of curriculum learning

  • We do not disagree with the existing work that shows no benefit from curricula with extended training in many cases. Figures 13 and 14 show that the effectiveness of our curriculum diminishes as IID reaches a sufficient budget to learn from an implicit curriculum. However, curricula do improve learning efficiency, which can be crucial in domains with a large number of training problems (such as our chess setting). Our goal in this work is to find an effective strategy for curriculum learning, especially compared to other curriculum strategies.
  • To extend our results to a more complex math domain with larger budgets, we run each curriculum on the Qwen3.5-0.8B model with GRPO on the DeepScaleR math dataset. Even with 300–500 steps of RL (6-10h of training on our hardware), we observe that LevelUp outperforms all baselines. We note that the model is still improving in all cases, and we expect IID to eventually catch up. However, this highlights that LevelUp is beneficial even with significant computational investment. Results and details are presented in Section R3.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.