Curriculum Learning as Transport: Understanding Curricula with Wasserstein Geodesics
Changho Shin, David Alvarez-Melis
cs.LG
Quick verdict
Transport lens decouples curriculum design factors; synthetic evidence is strong, but real-SFT gains vanish and headline comparisons lack significance tests.
Summary
This paper recasts curriculum learning as an optimal-transport problem: a curriculum is a path of sampling distributions over discrete difficulty levels, and Wasserstein geodesic interpolation between an easy-heavy and a hard-heavy endpoint decouples the design factors that are usually entangled — ordering, cumulative exposure, endpoint sharpness, pacing, and level geometry. On a synthetic suite of 12 tasks with 33 task-by-difficulty-axis conditions and three calibrated budgets (mostly small from-scratch transformers; two tasks fine-tune SmolLM-135M), the authors find that no strategy dominates; curricula mainly reallocate where a fixed budget pays off; and easy-to-hard ordering beats an exposure-matched static baseline on hardest-level accuracy, with forward beating reverse Wasserstein in 92 of 99 conditions — evidence that timing, not cumulative exposure alone, matters. A 9×9 sweep over endpoint temperature and pacing exponent shows very sharp endpoints are brittle. An adaptive-geometry prototype (WARP) adds only 0.9 points at medium budget, and the framework extends to graph-structured difficulty spaces. In real SFT on ARC, MMLU-STEM, and StrategyQA with three pretrained models, curriculum effects largely vanish and static sampling wins every block; a synthetic pretraining analogue suggests pretraining weakens these effects. Main limitations: synthetic-only main evidence, small untested differences, and a nonstandard exposure-adjusted accuracy metric.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
4 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Parameterizing curricula as Wasserstein paths over difficulty-level distributions decouples ordering, cumulative exposure, endpoint sharpness, pacing, and geometry within one controlled framework.
- No curriculum strategy dominates across 33 synthetic task-by-axis conditions and three budgets; curricula mainly reallocate where a fixed budget is spent most effectively (Table 1).
- Easy-to-hard ordering improves hardest-level performance beyond cumulative exposure: forward beats reverse Wasserstein in 92/99 conditions and beats exposure-matched static sampling on hard buckets.
- Pacing and endpoint smoothness substantially shape where a curriculum helps, with very sharp endpoints (τ=0.1) collapsing under both aggressive and conservative pacing (Fig. 4c).
- Pretraining weakens curriculum effects: static i.i.d. wins all nine real-SFT model-budget blocks, and a synthetic pretraining analogue shows gaps shrinking as pretraining grows (Finding 9).
Strengths
Clean conceptual decomposition of a confounded design space
Representing curricula as Wasserstein geodesic paths over difficulty-level distributions (Sec. 3.1) lets ordering, exposure, endpoints (τ), pacing (γ), and geometry vary one at a time — exactly the isolation prior work (Wu et al. 2021; Jia et al. 2026) lacked.
Rigorous controls for the exposure confound
The exposure-matched static baseline (Sec. 5) preserves cumulative per-level exposure while destroying temporal order, and reverse traversal preserves exposure while flipping direction; the finding that forward beats reverse in 92/99 conditions is a direct test that timing, not exposure alone, matters.
Unusually complete experimental specification
Honest negative results and mechanism probes
Weaknesses
No inferential statistics behind headline tallies
External validity of the main suite
Exposure-adjusted accuracy is a nonstandard, unvalidated metric
Extension results are prototype-grade
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This COLM 2026 paper (arXiv 2609.09099) proposes **Wasserstein curriculum paths**: a curriculum is a trajectory of sampling distributions over discrete difficulty levels, and training follows the W1 geodesic between an easy-heavy endpoint P0 and a hard-heavy endpoint P1, with endpoints (P1)ℓ ∝ exp(ℓ/τ) and pacing tk=(k/T)^γ (Sec. 3.1). The stated goal is analytical, not a new method: isolate ordering, cumulative exposure, endpoint smoothness, pacing, and geometry within one parameterization. Headline claims: (1) no curriculum dominates across 33 task-by-axis conditions and three budgets (Table 1); (2) curricula mainly reallocate where a fixed budget pays off (Finding 2); (3) easy-to-hard ordering beats exposure-matched static sampling on hardest-level accuracy — timing matters beyond exposure (Finding 3); (4) very sharp endpoints are brittle (Finding 4); (5) pretraining weakens curriculum effects (Finding 9). Extensions: WARP adaptive geometry (Sec. 7) and graph-structured difficulty spaces (Sec. 8).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 4 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Does curriculum learning actually work?
Per this paper, it depends strongly on context. Across 12 synthetic tasks, 33 difficulty settings, and three budgets, no curriculum strategy dominated. Curricula mainly reallocate where a fixed training budget pays off: easy-to-hard ordering helped hardest-level accuracy in the synthetic suite, but in fine-tuning pretrained models (ARC, MMLU-STEM, StrategyQA) static uniform sampling was the strongest baseline in every tested block.
What is a Wasserstein curriculum?
It represents a curriculum as a path of sampling distributions over ordered difficulty levels, moving between an easy-heavy and a hard-heavy endpoint via Wasserstein (optimal-transport) geodesic interpolation, so probability mass passes smoothly through intermediate difficulty levels instead of directly mixing the endpoint distributions as linear interpolation does.
Why do curriculum gains disappear in real LLM fine-tuning?
The authors' synthetic pretraining analogue shows curriculum effects shrink as pretraining grows: the easy-to-hard advantage fell from 13.5 to 2.3 points and the gap to exposure-matched static sampling from 3.9 to 0.2 points, suggesting that after pretraining there is less remaining room for improvement — one mechanism behind the weak effects observed in SFT.
Is code available for this paper?
The provided text contains no code or data release statement. The paper does give extensive per-task construction details (difficulty levels, budgets, architectures, examples) in its appendix, so reimplementation is feasible, but you would need to build the suite yourself unless the authors release code elsewhere.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The text is a sampled excerpt with three '[...sampling gap...]' regions (mid Sections 5–6, mid Appendix B.2, and the C.6–D.2 span plus middle references); statements about those regions rest on the surrounding visible text only.
- Figures appear only as extracted captions and axis labels; values plotted inside figures (e.g., the Fig. 4c τ×γ grid, Figs. 21/35–38) could not be independently verified.
- No code or data availability statement appears in the sampled text; one may exist in an unsampled portion.
- Appendix B.2 opens mid-sentence at a sampling gap ('All models were trained on a s…'); architecture details cited here come from the visible remainder of that subsection.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.