Elamur Research⌕
☰
home / papers / 2026 / Beyond Random Sampling: Efficient Language Model Pretraining via Curriculum Learning

Beyond Random Sampling: Efficient Language Model Pretraining via Curriculum Learning

Summary

Large language models are usually pretrained on documents fed in random order. This leaves open whether arranging data from simple to complex could save compute. The authors run what they present as the first systematic, large-scale test of curriculum learning for LLM pretraining. They use an English subset of CulturaX and decoder-only models based on the LLaMA3.2 architecture (0.5B to 3B parameters), training more than 200 models on budgets of up to 100B tokens.

Three strategies are compared:

  • strict easy-to-hard ordering;
  • sampling controlled by linear, quadratic and inverse quadratic pacing functions;
  • interleaved schedules.

Each strategy is driven by one of six cheap, model-independent difficulty scores. The experiments cover three settings: a fixed limited dataset, an effectively unlimited pool, and continued training.

Across eight benchmarks, curricula reach random-baseline quality noticeably sooner. A curriculum phase used as warmup before ordinary shuffled training gives gains that persist late into training. Since only the ordering changes, the method adds no training cost and can be combined with existing data filtering or mixture techniques.

Key points

  • The six difficulty metrics were chosen from 15 candidates via Spearman correlation analysis. The final set includes Perplexity and Number of Tokens, along with MTLD, Compression Ratio, Fertility and Flesch Reading Ease.
  • Curricula needed 18–45% fewer training steps to match the random-sampling baseline, mainly in early and mid training.
  • Curriculum-as-warmup improved scores by up to 3.5%. In the 100B-token run, the gain was about 3.1%.
  • Compression Ratio, MTLD and Flesch Reading Ease were the most useful signals. Perplexity performed poorly because its "hardest" samples, placed at the end of training, tend to be noisy.
  • The authors list several limitations: only decoder-only models and English data were tested, the difficulty scores are static and precomputed, and effects on downstream fine-tuning were not evaluated.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Linguistics & languages

Cited by

6 papers of our library cite it (of the 140673 whose bibliographies we read).

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.eacl-long.271 · arXiv 2506.11300

@inproceedings{zhang2026beyond,
  title = {{Beyond Random Sampling: Efficient Language Model Pretraining via Curriculum Learning}},
  author = {Zhang, Yang and Mohamed, Amr and Abdine, Hadi and Shang, Guokan and Vazirgiannis, Michalis},
  booktitle = {Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.eacl-long.271},
  eprint = {2506.11300},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.18653/v1/2026.eacl-long.271}
}

References

55 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774. find our card
  2. Mehdi Ali, Michael Fromm, Klaudia Thellmann, Richard Rutmann, Max Lübbering, Johannes Leveling, Katrin Klug, Jan Ebert, Niclas Doll, Jasper Buschhoff, et al. 2024. Tokenizer choice for llm training: Negligible or crucial? In Findings of the Association for Computational Linguistics: NAACL 2024, page… find
  3. Rohan Anil, Andrew M Dai, Orhan Firat, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. 2023. Palm 2 technical report. arXiv preprint arXiv:2305.10403. find our card
  4. Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. 2009a. Curriculum learning. In Proceedings of the 26th Annual International Conference on Machine Learning, ICML ’09, page 41–48, New York, NY, USA. Association for Computing Machinery. find our card
  5. Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. 2009b. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41–48. find our card
  6. Yonatan Bisk, Rowan Zellers, Jianfeng Gao, Yejin Choi, et al. 2020. Piqa: Reasoning about physical commonsense in natural language. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 7432–7439. find our card
  7. Mayee Chen, Nicholas Roberts, Kush Bhatia, Jue Wang, Ce Zhang, Frederic Sala, and Christopher Ré. 2023. Skill-it! a data-driven skills framework for understanding and training language models. Advances in Neural Information Processing Systems, 36:36000– 36040. find
  8. Volkan Cirik, Eduard Hovy, and Louis-Philippe Morency. 2016. Visualizing and understanding curriculum learning for long short-term memory networks. arXiv preprint arXiv:1611.06204. — [arXiv:1611.06204](https://arxiv.org/abs/1611.06204) find
  9. Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. 2019. Boolq: Exploring the surprising difficulty of natural yes/no questions. arXiv preprint arXiv:1905.10044. find our card
  10. Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. 2018. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv preprint arXiv:1803.05457. find
  11. Meri Coleman and Ta Lin Liau. 1975. A computer readability formula designed for machine scoring. Journal of Applied Psychology, 60(2):283. find
  12. Edgar Dale and Jeanne S Chall. 1948. A formula for predicting readability: Instructions. Educational research bulletin, pages 37–54. find
  13. Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac’h, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lintang Sutawika, Eric Tang, Ani… find
  14. Andrew Gordon, Zornitsa Kozareva, and Melissa Roemmele. 2012. SemEval-2012 task 7: Choice of plausible alternatives: An evaluation of commonsense causal reasoning. In *SEM 2012: The First Joint Conference on Lexical and Computational Semantics – Volume 1: Proceedings of the main conference and the s… find our card
  15. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783. find
  16. Robert Gunning. 1969. The fog index after twenty years. Journal of Business Communication, 6(2):3–13. find
  17. Guy Hacohen and Daphna Weinshall. 2019. On the power of curriculum learning in training deep networks. In International conference on machine learning, pages 2535–2544. PMLR. find
  18. Kenneth Heafield. 2011. KenLM: Faster and smaller language model queries. In Proceedings of the Sixth Workshop on Statistical Machine Translation, pages 187–197, Edinburgh, Scotland. Association for Computational Linguistics. find our card
  19. Yutong Hu, Quzhe Huang, Mingxu Tao, Chen Zhang, and Yansong Feng. 2024. Can perplexity reflect large language model’s ability in long text understanding? arXiv preprint arXiv:2405.06105. — [arXiv:2405.06105](https://arxiv.org/abs/2405.06105) find
  20. Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. 2020. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361. find
  21. Jisu Kim and Juhwan Lee. 2024. Strategic data ordering: Enhancing large language model performance through curriculum learning. arXiv preprint arXiv:2405.07490. — [arXiv:2405.07490](https://arxiv.org/abs/2405.07490) find
  22. J.P. Kincaid. 1975. Derivation of New Readability Formulas: (automated Readability Index, Fog Count and Flesch Reading Ease Formula) for Navy Enlisted Personnel. Research Branch report. Chief of Naval Technical Training, Naval Air Station Memphis. find
  23. M Kumar, Benjamin Packer, and Daphne Koller. 2010. Self-paced learning for latent variable models. Advances in neural information processing systems, 23. find
  24. Cao Liu, Shizhu He, Kang Liu, Jun Zhao, et al. 2018. Curriculum learning for natural answer generation. In IJCAI, pages 4223–4229. find
  25. Hong Liu, Sang Michael Xie, Zhiyuan Li, and Tengyu Ma. 2023. Same pre-training loss, better downstream: Implicit bias matters for language models. In International Conference on Machine Learning, pages 22188–22214. PMLR. find
  26. Shayne Longpre, Gregory Yauney, Emily Reif, Katherine Lee, Adam Roberts, Barret Zoph, Denny Zhou, Jason Wei, Kevin Robinson, David Mimno, et al. 2024. A pretrainer’s guide to training data: Measuring the effects of data age, domain coverage, quality, & toxicity. In Proceedings of the 2024 Conference… find
  27. Ilya Loshchilov and Frank Hutter. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101. find
  28. Max Marion, Ahmet Üstün, Luiza Pozzobon, Alex Wang, Marzieh Fadaee, and Sara Hooker. 2023. When less is more: Investigating data pruning for pretraining llms at scale. arXiv preprint arXiv:2309.04564. find
  29. G Harry Mc Laughlin. 1969. Smog grading-a new readability formula. Journal of reading, 12(8):639–646. find
  30. Bryan C McCannon. 2019. Readability and research impact. Economics Letters, 180:76–79. find
  31. Philip M McCarthy and Scott Jarvis. 2010. Mtld, vocd-d, and hd-d: A validation study of sophisticated approaches to lexical diversity assessment. Behavior research methods, 42(2):381–392. find
  32. Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. 2018. Can a suit of armor conduct electricity? a new dataset for open book question answering. arXiv preprint arXiv:1809.02789. find
  33. Brando Miranda, Alycia Lee, Sudharsan Sundar, Allison Casasola, and Sanmi Koyejo. 2023. Beyond scale: The diversity coefficient as a data quality metric for variability in natural language data. arXiv preprint arXiv:2306.13840. — [arXiv:2306.13840](https://arxiv.org/abs/2306.13840) find
  34. Koichi Nagatsuka, Clifford Broni-Bediako, and Masayasu Atsumi. 2023. Length-based curriculum learning for efficient pre-training of language models. New Generation Computing, 41(1):109–134. find
  35. Thuat Nguyen, Chien Van Nguyen, Viet Dac Lai, Hieu Man, Nghia Trung Ngo, Franck Dernoncourt, Ryan A. Rossi, and Thien Huu Nguyen. 2024. CulturaX: A cleaned, enormous, and multilingual dataset for large language models in 167 languages. In Proceedings of the 2024 Joint International Conference on Com… find our card
  36. Hadi Pouransari, Chun-Liang Li, Jen-Hao Chang, Pavan Kumar Anasosalu Vasu, Cem Koc, Vaishaal Shankar, and Oncel Tuzel. 2024. Dataset decomposition: Faster llm training with variable sequence length curriculum. Advances in Neural Information Processing Systems, 37:36121–36147. find
  37. Brian Richards. 1987. Type/token ratios: What do they really tell us? Journal of child language, 14(2):201– 209. find
  38. Noveen Sachdeva, Benjamin Coleman, Wang-Cheng Kang, Jianmo Ni, Lichan Hong, Ed H Chi, James Caverlee, Julian McAuley, and Derek Zhiyuan Cheng. 2024. How to train data-efficient llms. arXiv preprint arXiv:2402.09668. find
  39. Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. 2021. Winogrande: An adversarial winograd schema challenge at scale. Communications of the ACM, 64(9):99–106. find
  40. Samarth Sinha, Animesh Garg, and Hugo Larochelle. 2020. Curriculum by smoothing. Advances in Neural Information Processing Systems, 33:21653–21664. find
  41. Edgar A Smith and RJ Senter. 1967. Automated readability index, volume 66. Aerospace Medical Research Laboratories, Aerospace Medical Division, Air . . . . find
  42. Ben Sorscher, Robert Geirhos, Shashank Shekhar, Surya Ganguli, and Ari Morcos. 2022. Beyond neural scaling laws: beating power law scaling via data pruning. Advances in Neural Information Processing Systems, 35:19523–19536. find our card
  43. Petru Soviany, Radu Tudor Ionescu, Paolo Rota, and Nicu Sebe. 2022. Curriculum learning: A survey. International Journal of Computer Vision, 130(6):1526– 1565. find
  44. Valentin I Spitkovsky, Hiyan Alshawi, and Daniel Jurafsky. 2009. Baby steps: How “less is more” in unsupervised dependency parsing. NIPS: Grammar Induction, Representation of Language and Language Learning, pages 1–10. find
  45. Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, and Donald Metzler. 2021. Scale efficiently: Insights from pre-training and fine-tuning transformers. arXiv preprint arXiv:2109.10686. — [arXiv:2109.10686](https://arxiv.org/abs/2109.10686) find
  46. Alexey Tikhonov and Max Ryabinin. 2021. It’s all in the heads: Using attention heads as a baseline for cross-lingual transfer in commonsense reasoning. Preprint, arXiv:2106.12066. find
  47. Kushal Tirumala, Daniel Simig, Armen Aghajanyan, and Ari Morcos. 2023. D4: Improving llm pretraining via document de-duplication and diversification. Advances in Neural Information Processing Systems, 36:53983–53995. find
  48. Xiaoxia Wu, Ethan Dyer, and Behnam Neyshabur. 2020. When do curricula work? arXiv preprint arXiv:2012.03107. find
  49. Sang Michael Xie, Hieu Pham, Xuanyi Dong, Nan Du, Hanxiao Liu, Yifeng Lu, Percy S Liang, Quoc V Le, Tengyu Ma, and Adams Wei Yu. 2023. Doremi: Optimizing data mixtures speeds up language model pretraining. Advances in Neural Information Processing Systems, 36:69798–69818. find
  50. Yushi Yang, Andrew M Bean, Robert McCraith, and Adam Mahdi. 2024. Evaluating fine-tuning efficiency of human-inspired learning strategies in medical question answering. arXiv preprint arXiv:2408.07888. — [arXiv:2408.07888](https://arxiv.org/abs/2408.07888) find
  51. Mingjia Yin, Chuhan Wu, Yufei Wang, Hao Wang, Wei Guo, Yasheng Wang, Yong Liu, Ruiming Tang, Defu Lian, and Enhong Chen. 2024. Entropy law: The story behind data compression and llm performance. arXiv preprint arXiv:2407.06645. find
  52. Zichun Yu, Spandan Das, and Chenyan Xiong. 2024. Mates: Model-aware data selection for efficient pre-training with data influence models. Advances in Neural Information Processing Systems, 37:108735– 108759. find
  53. Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. 2019. Hellaswag: Can a machine really finish your sentence? arXiv preprint arXiv:1905.07830. find
  54. Xuan Zhang, Gaurav Kumar, Huda Khayrallah, Kenton Murray, Jeremy Gwinnup, Marianna J Martindale, Paul McNamee, Kevin Duh, and Marine Carpuat. 2018. An empirical exploration of curriculum learning for neural machine translation. arXiv preprint arXiv:1811.00739. — [arXiv:1811.00739](https://arxiv.org/abs/1811.00739) find
  55. Xuemiao Zhang, Liangyu Xu, Feiyu Duan, Yongwei Zhou, Sirui Wang, Rongxiang Weng, Jingang Wang, and Xunliang Cai. 2025. Preference curriculum: Llms should always be pretrained on their preferred data. arXiv preprint arXiv:2501.13126. find

↑ back to top