Elamur Research⌕
☰
home / papers / 2026 / Memorization, Emergence, and Explaining Reversal Failures: A Controlled Study of Relational Semantics in LLMs

Memorization, Emergence, and Explaining Reversal Failures: A Controlled Study of Relational Semantics in LLMs

Paper · 2026 · Yihua Zhu, Qianying Liu, Jiaxin Wang et al.

Summary

The paper asks whether autoregressive language models actually learn the logic behind relational words, such as symmetry (friend) and inverse pairs (father/son). It also asks whether the well-known reversal curse comes from missing inversion knowledge or from left-to-right generation order. Web-scale models mix polysemy with data contamination, so the authors build a fully controlled testbed instead. They sample synthetic knowledge-graph triples, render them as templated paragraphs, and train GPT2-style models from scratch, followed by QA fine-tuning. The models are then tested on three things: recalling seen facts, inferring missing symmetric or inverse facts, and in-context generalization to entities never seen in training.

The main findings:

  • Relational semantics appear abruptly, jumping from near-zero to near-perfect accuracy once there is enough logic-bearing training data. This happens even in 2–3 layer models.
  • Models that generalize show stable signals in their intermediate layers.
  • Forward and reverse queries compared under matched order, together with a diffusion-model baseline, suggest that reversal failures come mostly from the ordering bias of autoregressive generation rather than from a lack of inversion semantics.

This matters for interpreting the reversal curse and for choosing training-data designs, such as exposing models to both directions of a relation.

Key points

  • Synthetic corpus design:
    • Person names are built from three disjoint pools of 100 tokens each, giving up to 10^6 names.
    • Jobs are drawn from 300 real occupations.
    • Train graphs hold 7 triples, including both directions of each logic relation.
    • Evaluation and ICL graphs drop the reverse facts and use non-overlapping entities.
  • Verbalization and variables: each triple is rendered in one of four surface formats. The study controls template count K, training samples N, epochs E and depth L, while the evaluation set is fixed at 500 samples.
  • Evaluation tasks: Memorize QA, Logic QA, and ICL Logic QA. ICL is tested both by completion before SFT and by question answering after SFT.
  • A harder relational rule follows the same pattern: ICL completion accuracy stays near 0% at N=50,000, then rises to 47.6% at N=100,000. Symmetry and inversion reach about 95%.
  • Mitigation and limitations:
    • Bidirectional data exposure during training reduces reversal failures.
    • Diffusion LMs show little forward/reverse gap, though this result is only qualitative.
    • The extension to pretrained GPT-2 small covers ICL completion only.
    • Code: github.com/YihuaZhu111/KG_synthetic_LLM_train.

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, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 works

🧭 Topics 1Research subjects and languages the work is about.Language models & pretraining · 3,894 works

Cite

Cite as Yihua Zhu, Qianying Liu, Jiaxin Wang, Fei Cheng, Chaoran Liu, Akiko Aizawa, Sadao Kurohashi and Hidetoshi Shimodaira. 2026. Memorization, Emergence, and Explaining Reversal Failures: A Controlled Study of Relational Semantics in LLMs. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-long.1688

@inproceedings{zhu2026memorization,
  title = {{Memorization, Emergence, and Explaining Reversal Failures: A Controlled Study of Relational Semantics in LLMs}},
  author = {Yihua Zhu, Qianying Liu, Jiaxin Wang, Fei Cheng, Chaoran Liu, Akiko Aizawa, Sadao Kurohashi and Hidetoshi Shimodaira},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.1688},
  url = {https://doi.org/10.18653/v1/2026.acl-long.1688}
}

References

24 references, 13 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Zeyuan Allen-Zhu and Yuanzhi Li. 2023. Physics of language models: Part 3.1, knowledge storage and extraction. arXiv preprint arXiv:2309.14316.
  2. Nora Belrose, Zach Furman, Logan Smith, Danny Halawi, Igor Ostrovsky, Lev McKinney, Stella Biderman, and Jacob Steinhardt. 2023. Eliciting latent predictions from transformers with the tuned lens. arXiv preprint arXiv:2303.08112.
  3. Lukas Berglund, Meg Tong, Max Kaufmann, Mikita Balesni, Asa Cooper Stickland, Tomasz Korbak, and Owain Evans. 2023. The reversal curse: Llms trained on" a is b" fail to learn" b is a". arXiv preprint arXiv:2309.12288. find
  4. Ronen Eldan and Yuanzhi Li. 2023. Tinystories: How small can language models be and still speak coherent english? arXiv preprint arXiv:2305.07759.
  5. Nelson Elhage, Neel Nanda, Catherine Olsson, Tom Henighan, Nicholas Joseph, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, Nova DasSarma, Dawn Drain, Deep Ganguli, Zac Hatfield-Dodds, Danny Hernandez, Andy Jones, Jackson Kernion, Liane Lovitt, Kamal Ndousse, Dario Amodei, Tom Brown, Ja… find
  6. Zhengkai Lin, Zhihang Fu, Kai Liu, Liang Xie, Binbin Lin, Wenxiao Wang, Deng Cai, Yue Wu, and Jieping Ye. 2024. Delving into the reversal curse: How far can large language models generalize? Advances in Neural Information Processing Systems, 37:30686– 30726. find
  7. Mingjie Liu, Shizhe Diao, Ximing Lu, Jian Hu, Xin Dong, Yejin Choi, Jan Kautz, and Yi Dong. 2025. Prorl: Prolonged reinforcement learning expands reasoning boundaries in large language models. arXiv preprint arXiv:2505.24864.
  8. Ang Lv, Kaiyi Zhang, Shufang Xie, Quan Tu, Yuhan Chen, Ji-Rong Wen, and Rui Yan. 2024. An analysis and mitigation of the reversal curse. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 13603–13615. find
  9. Neel Nanda, Lawrence Chan, Tom Lieberum, Jess Smith, and Jacob Steinhardt. 2023. Progress measures for grokking via mechanistic interpretability. arXiv preprint arXiv:2301.05217.
  10. Shen Nie, Fengqi Zhu, Zebin You, Xiaolu Zhang, Jingyang Ou, Jun Hu, Jun Zhou, Yankai Lin, Ji-Rong Wen, and Chongxuan Li. 2025. Large language diffusion models. In The Thirty-ninth Annual Conference on Neural Information Processing Systems. nostalgebraist. 2020. interpreting gpt: the logit lens. Less…
  11. Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730–27744.
  12. Shirui Pan, Linhao Luo, Yufei Wang, Chen Chen, Jiapu Wang, and Xindong Wu. 2024. Unifying large language models and knowledge graphs: A roadmap. IEEE Transactions on Knowledge and Data Engineering, 36(7):3580–3599.
  13. Alethea Power, Yuri Burda, Harri Edwards, Igor Babuschkin, and Vedant Misra. 2022. Grokking: Generalization beyond overfitting on small algorithmic datasets. arXiv preprint arXiv:2201.02177.
  14. Boshi Wang, Xiang Yue, Yu Su, and Huan Sun. 2024. Grokked transformers are implicit reasoners: A mechanistic journey to the edge of generalization. arXiv preprint arXiv:2405.15071. find
  15. Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. 2022a. Emergent abilities of large language models. arXiv preprint arXiv:2206.07682.
  16. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022b. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837.
  17. Qinyuan Wu, Soumi Das, Mahsa Amani, Bishwamittra Ghosh, Mohammad Aflah Khan, Krishna P Gummadi, and Muhammad Bilal Zafar. 2025. Rote learning considered useful: Generalizing over memorized data in llms. arXiv preprint arXiv:2507.21914. — [arXiv:2507.21914](https://arxiv.org/abs/2507.21914) find
  18. Lifan Yuan, Weize Chen, Yuchen Zhang, Ganqu Cui, Hanbin Wang, Ziming You, Ning Ding, Zhiyuan Liu, Maosong Sun, and Hao Peng. 2025. From f (x) and g(x) to f (g(x)): Llms learn new skills in rl by composing old ones. arXiv preprint arXiv:2509.25123. — [arXiv:2509.25123](https://arxiv.org/abs/2509.25123) find
  19. Charlie Zhang, Graham Neubig, and Xiang Yue. 2025. On the interplay of pre-training, mid-training, and rl on reasoning language models. arXiv preprint arXiv:2512.07783. — [arXiv:2512.07783](https://arxiv.org/abs/2512.07783) find
  20. Hanlin Zhu, Baihe Huang, Shaolun Zhang, Michael Jordan, Jiantao Jiao, Yuandong Tian, and Stuart J Russell. 2024. Towards a theoretical understanding of the’reversal curse’via training dynamics. Advances in Neural Information Processing Systems, 37:90473– 90513. find
  21. Yihua Zhu and Hidetoshi Shimodaira. 2024. 3D rotation and translation for hyperbolic knowledge graph embedding. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1497–1515, St. Julian’s, Malta. Association fo… find
  22. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9. find
  23. Zhiqing Sun, Zhi-Hong Deng, Jian-Yun Nie, and Jian Tang. 2019. Rotate: Knowledge graph embedding by relational rotation in complex space. arXiv preprint arXiv:1902.10197.
  24. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhu-2026-memorization-emergence-explaining-reversal-failures/