Elamur Research⌕
☰
home / papers / 2026 / Enhancing Multimodal Large Language Models for Ancient Chinese Character Evolution Analysis via Glyph-Driven Fine-Tuning

Enhancing Multimodal Large Language Models for Ancient Chinese Character Evolution Analysis via Glyph-Driven Fine-Tuning

Paper · 2026 · Rui Song, Lida Shi, Ruihua Qi et al.

Summary

Large multimodal models are increasingly applied to ancient Chinese scripts, but no one has systematically tested whether they can follow how individual characters changed shape over time. The authors addressed this gap. They merged existing evolution datasets with tracings from Vividict into a facsimile-level collection of 7,740 characters and nearly 30,000 glyph images. These span five stages, from Oracle Bone Script to Regular Script. With input from paleographers, they built a benchmark of 11 subtasks in three groups, totalling more than 130,000 questions. They tested 19 MLLMs on it. The models handled style and glyph comparison reasonably well but performed poorly at recognizing characters and reasoning about evolution. The authors then proposed GEVO, a contrastive fine-tuning scheme that is glyph-driven and ordered in a curriculum-like way. A 2B model tuned with GEVO improved on every task, outperformed all evaluated models, and generalized to out-of-distribution data. The released benchmark and models give the field a reusable testbed for this problem.

Key points

  • Benchmark: 3 task categories with 11 subtasks and over 130k instances.
    • Tasks are posed as question answering, binary judgment, or multiple choice.
    • Each task is split 9:1 into training and test sets.
    • Instructions were drafted with ChatGPT and then checked by experts.
    • Oracle bone characters form the smallest portion of the data, since only around two thousand have been reliably deciphered.
  • Evaluation: 19 MLLMs ranging from 1B to 72B parameters, including only three closed-source models because of cost.
    • The closed-source models often refused to answer or recognized almost no characters.
    • Placing tasks inside an explicit evolutionary context usually improved results.
  • GEVO results: fine-tuning was built on Qwen3-VL-2B-Instruct.
    • Recognition accuracy rose above 50% for clerical and regular script.
    • Gains for oracle bone and bronze scripts stayed small.
  • OOD test: the authors built a subset from OBIsEvolution with 148 characters and 717 facsimiles.
    • GEVO averaged about 81.8, compared with roughly 43.1 for Qwen3-VL-2B and 61.9 for Qwen3-VL-8B.
  • Limitations and resources:
    • Only clean hand-copied facsimiles were used, not ink rubbings.
    • Code, data, and models are available at github.com/songruiecho/GEVO.

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

🎓Jilin UniversityChangchun · China25 works · 5,252 citations

 Fields, topics and concepts

📚 Shelves 5The 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 worksComputer Vision and Pattern Recognition · 2,869 works

🧭 Topics 2Research subjects and languages the work is about.Language models & pretraining · 3,320 worksVision & language · 657 works

Cite

Cite as Rui Song, Lida Shi, Ruihua Qi, Yingji Li and Hao Xu. 2026. Enhancing Multimodal Large Language Models for Ancient Chinese Character Evolution Analysis via Glyph-Driven Fine-Tuning. 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.1043

@inproceedings{song2026enhancing,
  title = {{Enhancing Multimodal Large Language Models for Ancient Chinese Character Evolution Analysis via Glyph-Driven Fine-Tuning}},
  author = {Rui Song, Lida Shi, Ruihua Qi, Yingji Li and Hao Xu},
  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.1043},
  url = {https://doi.org/10.18653/v1/2026.acl-long.1043}
}

References

32 references, 9 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Roar Bökset. 2006. Long story of short forms: the evolution of simplified Chinese characters. Ph.D. thesis, Institutionen för orientaliska språk. find
  2. Jiahuan Cao, Yang Liu, Peirong Zhang, Yongxin Shi, Kai Ding, and Lianwen Jin. 2025. Tonggu-vl: Advancing visual-language understanding in chinese classical studies through parameter sensitivity-guided instruction tuning. In Proceedings of the 33rd ACM International Conference on Multimedia, pages 11… — [arXiv:2510.26114](https://arxiv.org/abs/2510.26114) find
  3. Jing Li, Xueke Chi, Qiufeng Wang, Kaizhu Huang, Da-Han Wang, Yongge Liu, and Cheng-Lin Liu. 2026. A comprehensive survey of oracle character recognition: Challenges, datasets, methodology, and beyond. Pattern Recognition, 169:111824. find
  4. Zijian Chen, Wenjun Zhang, Guangtao Zhai, and 1 others. 2025. Obi-bench: Can lmms aid in study of ancient script on oracle bones? In The Thirteenth International Conference on Learning Representations. find
  5. Jing Li, Bin Dong, Qiu-Feng Wang, Lei Ding, Rui Zhang, and Kaizhu Huang. 2023a. Decoupled learning for long-tailed oracle character recognition. In International Conference on Document Analysis and Recognition, pages 165–181. find
  6. Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, and 1 others. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arX…
  7. Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, and 1 others. 2024. Chatglm: A family of large language models from glm-130b to glm-4 all tools. arXiv preprint arXiv:2406.12793.
  8. Haisu Guan, Jinpeng Wan, Yuliang Liu, Pengjie Wang, Kaile Zhang, Zhebin Kuang, Xinyu Wang, Xiang Bai, and Lianwen Jin. 2024a. An open dataset for the evolution of oracle bone characters: EVOBC. CoRR, abs/2401.12467. — [arXiv:2401.12467](https://arxiv.org/abs/2401.12467) find
  9. Haisu Guan, Huanxin Yang, Xinyu Wang, Shengwei Han, Yongge Liu, Lianwen Jin, Xiang Bai, and Yuliang Liu. 2024b. Deciphering oracle bone language with diffusion models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangko… find
  10. Shuangping Huang, Haobin Wang, Yongge Liu, Xiaosong Shi, and Lianwen Jin. 2019. Obc306: A large-scale oracle bone character recognition dataset. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 681–688. find
  11. Jing Li, Qiu-Feng Wang, Kaizhu Huang, Xi Yang, Rui Zhang, and John Y Goulermas. 2023b. Towards better long-tailed oracle character recognition with adversarial data augmentation. Pattern Recognition, 140:109534. find
  12. Jinhao Li, Zijian Chen, Runze Jiang, Tingzhu Chen, Changbo Wang, and Guangtao Zhai. 2025b. Mitigating long-tail distribution in oracle bone inscriptions: Dataset, model, and benchmark. In Proceedings of the 33rd ACM International Conference on Multimedia, pages 7729–7738. find
  13. Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2023. Improved baselines with visual instruction tuning. Preprint, arXiv:2310.03744.
  14. Yang Liu, Jiahuan Cao, Hiuyi Cheng, Yongxin Shi, Kai Ding, and Lianwen Jin. 2025. Mcs-bench: A comprehensive benchmark for evaluating multimodal large language models in chinese classical studies. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: L… find
  15. Runqi Qiao, Qiuna Tan, Guanting Dong, MinhuiWu MinhuiWu, Jiapeng Wang, Yifan Zhang, Zhuoma GongQue, Chong Sun, Yida Xu, Yadong Xue, and 1 others. 2025. V-oracle: Making progressive reasoning in deciphering oracle bones for you and me. In Proceedings of the 63rd Annual Meeting of the Association for… find
  16. Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, and 1 others. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276.
  17. Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, and 1 others. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748–8…
  18. Qingju Jiao, Jingwen Wu, Qi Liu, Han Zhang, Zhan Zhang, Bang Li, Jing Xiong, Guoying Liu, and Yongge Liu. 2025. A graph-based evolutionary dataset for oracle bone characters from inscriptions to modern chinese scripts. npj Heritage Science, 13(1):369. find
  19. Daqian Shi, Xiaolei Diao, Lida Shi, Hao Tang, Yang Chi, Chuntao Li, and Hao Xu. 2022. Charformer: A glyph fusion based attentive framework for high-precision character image denoising. In Proceedings of the 30th ACM international conference on multimedia, pages 1147–1155. find
  20. Caoshuo Li, Zengmao Ding, Xiaobin Hu, Bang Li, Donghao Luo, Xu Peng, Taisong Jin, Yongge Liu, Shengwei Han, Jing Yang, and 1 others. 2025a. Oracleagent: A multimodal reasoning agent Qwen Team. 2025a. Qwen2.5-vl. find
  21. Qwen Team. 2025b. Qwen3 technical report. Preprint, arXiv:2505.09388.
  22. Mei Wang, Weihong Deng, and Sen Su. 2024a. Oracle character recognition using unsupervised discriminative consistency network. Pattern Recognition, 148:110180. find
  23. Mengru Wang, Yu Cai, Li Gao, Ruichen Feng, Qingju Jiao, Xiaolin Ma, and Yu Jia. 2022. Study on the evolution of chinese characters based on few-shot learning: From oracle bone inscriptions to regular script. Plos one, 17(8):e0272974. find
  24. Pengjie Wang, Kaile Zhang, Xinyu Wang, Shengwei Han, Yongge Liu, Jinpeng Wan, Haisu Guan, Zhebin Kuang, Lianwen Jin, Xiang Bai, and 1 others. 2024b. An open dataset for oracle bone character recognition and decipherment. Scientific Data, 11(1):976. find
  25. Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, Long Cui, Xingguang Wei, Zhaoyang Liu, Linglin Jing, Shenglong Ye, Jie Shao, and 1 others. 2025. Internvl3.5: Advancing open-source multimodal models in versatility, reasoning, and efficiency. arXiv preprint arXiv:2508.18265.
  26. Xin Wang, Yudong Chen, and Wenwu Zhu. 2021. A survey on curriculum learning. IEEE transactions on pattern analysis and machine intelligence, 44(9):4555–4576. find
  27. Haoran Wei, Yaofeng Sun, and Yukun Li. 2025. Deepseek-ocr: Contexts optical compression. arXiv preprint arXiv:2510.18234.
  28. Xing Wen. 2011. Chinese paleography, calligraphy, and pattern recognition: Styles and scripts in excavated ancient chinese documents. In 2011 International Conference on Document Analysis and Recognition, pages 951–956. find
  29. Xinyu Yao, Mengdi Wang, Bo Chen, and Xiaobing Zhao. 2025. Wenyangpt: A large language model for classical chinese tasks. arXiv preprint arXiv:2504.20609. — [arXiv:2504.20609](https://arxiv.org/abs/2504.20609) find
  30. Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, and 1 others. 2024. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800.
  31. Xinyi Zhao, Siyuan Liu, Yikai Wang, and Yanwei Fu. 2022. Ffd augmentor: towards few-shot oracle character recognition from scratch. In Proceedings of the Asian Conference on Computer Vision, pages 1622– 1639. find
  32. Zhihan Zhou, Daqian Shi, Rui Song, Lida Shi, Xiaolei Diao, and Hao Xu. 2025. Ancientbench: Towards comprehensive evaluation on excavated and transmitted chinese corpora. arXiv preprint arXiv:2512.17756. — [arXiv:2512.17756](https://arxiv.org/abs/2512.17756) find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/song-2026-enhancing-multimodal-large-language-models/