Elamur Research⌕
☰
home / papers / 2026 / Annotation Entropy Predicts Per-Example Learning Dynamics in LoRA Fine-Tuning

Annotation Entropy Predicts Per-Example Learning Dynamics in LoRA Fine-Tuning

Paper · 2026 · Brady Steele

Summary

Standard fine-tuning treats every training example as equally learnable, even though human annotators often disagree on NLI items. This student-workshop paper tests whether an external, model-independent measure of ambiguity can predict how individual examples are learned under LoRA. The author computes annotation entropy from ChaosNLI's 100 labels per item. Per-example loss against the majority label is logged at 39 checkpoints and summarized as area under the loss curve (AULC) and as net loss change. The experiments cover six models (four encoders, two decoder-only Qwen models) on SNLI and MNLI, several LoRA ranks, and full fine-tuning and IA3 baselines. Across all 25 conditions, items with higher entropy are learned more slowly. The more notable finding is that under LoRA the most contested items get worse over training, with their loss rising. In encoders, this behaviour mostly disappears under full fine-tuning or IA3. This matters for practitioners because low-rank adapters may quietly give up on genuinely ambiguous data. Monitoring per-example loss can reveal when this happens.

Key points

  • Data and binning. ChaosNLI subsets are used: SNLI (~1,514 items) and MNLI-matched (~1,599). Each is mixed with 20,000 bulk training examples. Entropy bins are clean (<0.4 nats), ambiguous (0.4–0.7) and contested (≥0.7). On the SNLI subset this gives 25.0% clean, 50.6% ambiguous and 24.4% contested.
  • Models and setup. The models are RoBERTa-base, BERT-base, DistilBERT, DeBERTa v3-base, Qwen2.5-1.5B and Qwen2.5-3B. LoRA uses r ∈ {4, 16}, with a sweep over r = 1–32 for RoBERTa and BERT on SNLI. Every configuration is run with 3 seeds.
  • Correlation results.
    • Spearman ρ between AULC and entropy is positive everywhere, ranging from 0.06 to 0.43. It is significant in 21 of 25 conditions after Bonferroni correction.
    • Decoder-only models reach ρ = 0.35–0.42, stronger than encoders at the same rank.
    • The correlation grows monotonically with LoRA rank.
  • Unlearning under LoRA. LoRA-specific unlearning of contested items persists even with soft-label (full annotator distribution) training. It also survives partial-correlation controls. MNLI shows weaker effects (ρ = 0.06–0.20).
  • Noise injection. A synthetic noise-injection test shows a small but significant effect (Cohen's d = 0.20).
  • Availability. Code and per-example tracking data are at github.com/BradySteele/lora-annotation-dynamics.

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

 Institutions

🎓Georgia Institute of TechnologyAtlanta · United States133 works · 24,497 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

Cite

Cite as Brady Steele. 2026. Annotation Entropy Predicts Per-Example Learning Dynamics in LoRA Fine-Tuning. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-srw.11

@inproceedings{steele2026annotation,
  title = {{Annotation Entropy Predicts Per-Example Learning Dynamics in LoRA Fine-Tuning}},
  author = {Brady Steele},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-srw.11},
  url = {https://doi.org/10.18653/v1/2026.acl-srw.11}
}

References

48 references, 18 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Devansh Arpit, Stanisław Jastrz˛ebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S. Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, and Simon Lacoste-Julien. 2017. A closer look at memorization in deep networks. In Proceedings of the International Conference on Ma… find
  2. Valerio Basile, Michael Fell, Tommaso Fornaciari, Dirk Hovy, Silviu Paun, Barbara Plank, Massimo Poesio, and Alexandra Uma. 2021. We need to consider disagreement in evaluation. In Proceedings of the 1st Workshop on Benchmarking: Past, Present and Future (BPPF).
  3. Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. 2009. Curriculum learning. In Proceedings of the International Conference on Machine Learning (ICML).
  4. Dan Biderman, Jacob Portes, Jose Javier Gonzalez Ortiz, Mansheej Paul, Philip Greengard, Connor Jennings, Daniel King, Sam Havens, Vitaliy Chiley, Jonathan Frankle, Cody Blakeney, and John P. Cunningham. 2024. LoRA learns less and forgets less. Transactions on Machine Learning Research (TMLR). find
  5. Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015. A large annotated corpus for learning natural language inference. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP).
  6. Beiduo Chen, Tiancheng Hu, Caiqi Zhang, Robert Litschko, Anna Korhonen, and Barbara Plank. 2026. Decoupling the effect of chain-of-thought reasoning: A human label variation perspective. arXiv preprint arXiv:2601.03154. find
  7. Aida Mostafazadeh Davani, Mark Díaz, and Vinodkumar Prabhakaran. 2022. Dealing with disagreements: Looking beyond the majority vote in subjective annotations. Transactions of the Association for Computational Linguistics (TACL), 10:92–110.
  8. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics (NAACL).
  9. Ning Ding, Yujia Qin, Guang Yang, Fuchao Wei, Zonghan Yang, Yusheng Su, Shengding Hu, Yulin Chen, Chi-Min Chan, Weizhu Chen, Jian Yi, Weilin Zhao, Xiaozhi Wang, Zhiyuan Liu, Hai-Tao Zheng, Jianfei Chen, Yang Liu, Jie Tang, Juanzi Li, and Maosong Sun. 2023. Parameter-efficient fine-tuning of large-sc…
  10. Abinitha Gourabathina, Hyewon Jeong, Teya Bergamaschi, Marzyeh Ghassemi, and Collin M. Stultz. 2026. Robustness beyond known groups with low-rank adaptation. arXiv preprint arXiv:2602.06924. — [arXiv:2602.06924](https://arxiv.org/abs/2602.06924) find
  11. Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel R. Bowman, and Noah A. Smith. 2018. Annotation artifacts in natural language inference data. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics (NAACL).
  12. Guy Hacohen, Leshem Choshen, and Daphna Weinshall. 2020. Let’s agree to agree: Neural networks share classification order on real datasets. In Proceedings of the 37th International Conference on Machine Learning (ICML). find
  13. Pengcheng He, Jianfeng Gao, and Weizhu Chen. 2023. DeBERTaV3: Improving DeBERTa using ELECTRA-style pre-training with gradient-disentangled embedding sharing. In Proceedings of the International Conference on Learning Representations (ICLR).
  14. Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-rank adaptation of large language models. In Proceedings of the International Conference on Learning Representations (ICLR).
  15. Sangyoon Lee and Jaeho Lee. 2026. Beware of the batch size: Hyperparameter bias in evaluating LoRA. arXiv preprint arXiv:2602.09492. — [arXiv:2602.09492](https://arxiv.org/abs/2602.09492) find
  16. Seohyeong Lee, Eunwon Kim, Hwaran Lee, and Buru Chang. 2025. Dataset cartography for large language model alignment: Mapping and diagnosing preference data. arXiv preprint arXiv:2505.23114. — [arXiv:2505.23114](https://arxiv.org/abs/2505.23114) find
  17. Elisa Leonardelli, Gavin Abercrombie, Dina Almanea, Valerio Basile, Tommaso Fornaciari, Barbara Plank, Verena Rieser, Alexandra Uma, and Massimo Poesio. 2023. SemEval-2023 task 11: Learning with disagreements (LeWiDi). In Proceedings of the 17th International Workshop on Semantic Evaluation (SemEval…
  18. Elisa Leonardelli, Silvia Casola, Siyao Peng, and 1 others. 2025. LeWiDi-2025 at NLPerspectives: Third edition of the learning with disagreements shared task. arXiv preprint arXiv:2510.08460. find
  19. Tom McCoy, Ellie Pavlick, and Tal Linzen. 2019. Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference. In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL).
  20. Johannes Mario Meissner, Napat Thumwanit, Saku Sugawara, and Akiko Aizawa. 2021. Embracing ambiguity: Shifting the training target of NLI models. In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL).
  21. Soren Mindermann, Jan M. Brauner, Muhammed T. Razzak, Mrinank Sharma, Andreas Kirsch, Winnie Xu, Benedikt Höltgen, Aidan N. Gomez, Adrien Morisot, Sebastian Farquhar, and Yarin Gal. 2022. Prioritized training on points that are learnable, worth learning, and not yet learnt. In Proceedings of the 39t… find
  22. Andrei Mircea, Supriyo Chakraborty, Nima Chitsazan, Milind Naphade, Sambit Sahu, Irina Rish, and Ekaterina Lobacheva. 2025. Training dynamics underlying language model scaling laws: Loss deceleration and zero-sum learning. In Proceedings of the Annual Meeting of the Association for Computational Lin… find
  23. Yixin Nie, Xiang Zhou, and Mohit Bansal. 2020. What can we learn from collective human opinions on natural language inference data? In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP).
  24. Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL). find
  25. Mansheej Paul, Surya Ganguli, and Gintare Karolina Dziugaite. 2021. Deep learning on a data diet: Finding important examples early in training. In Advances in Neural Information Processing Systems (NeurIPS). find
  26. Haokun Liu, Derek Tam, Mohammed Muqeeth, Jay Mohta, Tenghao Huang, Mohit Bansal, and Colin Raffel. 2022. Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning. In Advances in Neural Information Processing Systems (NeurIPS).
  27. Ellie Pavlick and Tom Kwiatkowski. 2019. Inherent disagreements in human textual inferences. Transactions of the Association for Computational Linguistics (TACL), 7:677–694.
  28. Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. 2024. DoRA: Weight-decomposed low-rank adaptation. In Proceedings of the International Conference on Machine Learning (ICML). find
  29. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. RoBERTa: A robustly optimized BERT pretraining approach. arXiv preprint arXiv:1907.11692.
  30. Paul K. Mandal. 2025. When is dataset cartography ineffective? using training dynamics does not improve robustness against adversarial SQuAD. arXiv preprint arXiv:2503.18290. — [arXiv:2503.18290](https://arxiv.org/abs/2503.18290) find
  31. Barbara Plank. 2022. The “problem” of human label variation: On ground truth in data, modeling and evaluation. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP).
  32. Geoff Pleiss, Tianyi Zhang, Ethan R. Elenberg, and Kilian Q. Weinberger. 2020. Identifying mislabeled data using the area under the margin ranking. In Advances in Neural Information Processing Systems (NeurIPS). find
  33. Zhenting Qi, Fan Nie, Alexandre Alahi, James Zou, Himabindu Lakkaraju, Yilun Du, Eric Xing, Sham Kakade, and Hanlin Zhang. 2025. EvoLM: In search of lost language model training dynamics. In Advances in Neural Information Processing Systems (NeurIPS). Oral. find
  34. Zahra Rahimi Afzal, Tara Esmaeilbeig, Mojtaba Soltanalian, and Mesrob I. Ohannessian. 2025. Linearization explains fine-tuning in large language models. In Advances in Neural Information Processing Systems (NeurIPS). find
  35. Soheil Zibakhsh Shabgahi, Pedram Aghazadeh, and Farinaz Koushanfar. 2026. Beyond perplexity: A lightweight benchmark for knowledge retention in supervised fine-tuning. arXiv preprint arXiv:2601.03505. — [arXiv:2601.03505](https://arxiv.org/abs/2601.03505) find
  36. Darshita Rathore, Vineet Kumar, Chetna Bansal, and Anindya Moitra. 2025. How much is too much? exploring LoRA rank trade-offs for retaining knowledge and domain robustness. In Proceedings of the International Joint Conference on Natural Language Processing (AACL-IJCNLP). find
  37. A Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. 2019. DistilBERT, a distilled version of BERT: Smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108. find
  38. Per-Seed Breakdown Table 3 reports the per-seed AULC–entropy correlation for RoBERTa on SNLI, showing the individual seed values underlying the means reported in Table 1. Table 4 reports the corresponding per-seed breakdown for the decoder-only models. Method Seed AULC ρ LoRA r=4 find
  39. Joanna Sliwa, Frank Schneider, Philipp Hennig, and Jose Miguel Hernandez-Lobato. 2025. Mitigating forgetting in low rank adaptation (LaLoRA). arXiv preprint arXiv:2512.17720. — [arXiv:2512.17720](https://arxiv.org/abs/2512.17720) find
  40. Swabha Swayamdipta, Roy Schwartz, Nicholas Lourie, Yizhong Wang, Hannaneh Hajishirzi, Noah A. Smith, and Yejin Choi. 2020. Dataset cartography: Mapping and diagnosing datasets with training dynamics. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP). find
  41. Qwen Team. 2024. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115. find
  42. Mariya Toneva, Alessandro Sordoni, Remi Tachet des Combes, Adam Trischler, Yoshua Bengio, and Geoffrey J. Gordon. 2019. An empirical study of example forgetting during deep neural network learning. In Proceedings of the International Conference on Learning Representations (ICLR). find
  43. Wanru Zhao, Hongxiang Fan, Shell Xu Hu, Wangchunshu Zhou, Bofan Chen, and Nicholas D. Lane. 2024. CLUES: Collaborative high-quality data selection for LLMs via training dynamics. In Advances in Neural Information Processing Systems (NeurIPS). 42 123 456 0.304 0.327 0.295 3.1×10 1.7×10 −31 8.8×10 −26… find
  44. LoRA r=16 42 123 456 0.330 0.366 0.341 4.3×10 −32 1.1×10 −39 2.0×10 −34 0.607 0.640 0.624 Full FT 0.416 0.444 0.419 5.9×10 −52 1.0×10 −59 1.5×10 −52 0.667 0.693 0.677 Model Qwen 1.5B Adina Williams, Nikita Nangia, and Samuel R. Bowman. 2018. A broad-coverage challenge corpus for sentence understandi…
  45. Xiao Zhang and Ji Wu. 2024. Dissecting learning and forgetting in language model finetuning. In Proceedings of the International Conference on Learning Representations (ICLR). find
  46. −27 42 123 456 *Table 3: Per-seed AULC–entropy Spearman ρ for RoBERTa on SNLI. All seeds yield highly significant correlations (p \< 10 −25 ). The r=4 \< r=16 \< Full FT ordering holds within every seed. Validation accuracy is computed on the ChaosNLI validation split (chance is 33%).* Alexandra Uma, T… find
  47. Yuchen Zeng and Kangwook Lee. 2024. The expressive power of low-rank adaptation. In Proceedings of the 12th International Conference on Learning Representations (ICLR). find
  48. Val Acc p-value Qwen 3B Method Seed LoRA r=4 42 123 456 0.354 0.243 0.365 0.252 0.325 0.220 0.700 0.716 0.660 LoRA r=16 42 123 456 0.397 0.273 0.409 0.283 0.363 0.245 0.703 0.729 0.693 LoRA r=4 42 123 456 0.392 0.266 0.385 0.262 0.374 0.259 0.723 0.733 0.749 LoRA r=16 42 123 0.417 0.282 0.428 0.294… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/steele-2026-annotation-entropy-predicts-per-example/