Elamur Research⌕
☰
home / papers / 2026 / CSPO: Alleviating Reward Ambiguity for Structured Table-to-LaTeX Generation

CSPO: Alleviating Reward Ambiguity for Structured Table-to-LaTeX Generation

Paper · 2026 · Yunfan Yang, Cuiling Lan, Jitao Sang et al.

Summary

Turning images of tables into LaTeX code is useful for digitizing scientific documents. However, multimodal LLMs often get cell merges, line styles or cell values wrong, or produce code that does not compile. The authors argue that standard RL post-training (e.g., GRPO) scores the whole output with one reward. This mixes structure, content and style into a single signal, which they call "reward ambiguity". As a result, a correct part of the code may be penalized while a broken part is reinforced.

They propose Component-Specific Policy Optimization (CSPO). A LaTeX parser splits each generated sequence into functional parts: packages, structure, cell appearance, caption, column alignment and line style. Each part is scored separately, with an LLM judge, and its advantage updates only the tokens belonging to that part. A global reward based on TEDS and compile status is kept alongside.

The authors also define hierarchical metrics for structure, content and style fidelity. Using Qwen2.5-VL-3B/7B trained on arXiv data, CSPO beats SFT and GRPO both in-domain and out-of-domain. This suggests that component-wise credit assignment matters for structured generation.

Key points

  • 3B results (GPT-4o judge): Overall Fidelity rises from 3.1% for the base Qwen2.5-VL-3B to 39.8% with SFT, 42.0% with GRPO and 45.2% with CSPO.
  • Out-of-domain results: on DocGenome-table-1k, CSPO improves Overall Fidelity over GRPO by 2.1% (3B) and 2.3% (7B). It also outperforms GRPO/VSGRPO on Table2LaTeX-test-simple.
  • Ablations:
    • Simply summing the component rewards into one scalar (CSPO w/ Comp. Sum) scores about 3% lower in Overall Fidelity than full CSPO.
    • Removing the global reward costs about 0.5%.
    • Structure and style rewards matter more than the content reward.
  • Judge reliability: results were checked with GPT-4o, Qwen3-Next-80B, DeepSeek-v3.2 and GPT-5.2 as judges. GPT-4o agreed with humans on roughly 90% of 500 sampled cases.
  • Limitations: the training set is small (15,000 samples from permissively licensed arXiv papers). The best 7B model reaches only about 53% Overall Fidelity, and LLM-based rewards add training cost.

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

🏢Microsoft Research AsiaBeijing · China82 works · 61,484 citations🎓Beijing Jiaotong UniversityBeijing · China28 works · 1,754 citations

 Fields, topics and concepts

📚 Shelves 2The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksComputer Vision and Pattern Recognition · 3,486 works

Cite

Cite as Yunfan Yang, Cuiling Lan, Jitao Sang and Yan Lu. 2026. CSPO: Alleviating Reward Ambiguity for Structured Table-to-LaTeX Generation. 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.1163

@inproceedings{yang2026cspo,
  title = {{CSPO: Alleviating Reward Ambiguity for Structured Table-to-LaTeX Generation}},
  author = {Yunfan Yang, Cuiling Lan, Jitao Sang and Yan Lu},
  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.1163},
  url = {https://doi.org/10.18653/v1/2026.acl-long.1163}
}

References

24 references, 11 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, and 1 others. 2025. Qwen2.5-vl technical report. arXiv preprint arXiv:2502.13923.
  2. Lukas Blecher, Guillem Cucurull, Thomas Scialom, and Robert Stojnic. 2023. Nougat: Neural optical understanding for academic documents. arXiv preprint arXiv:2308.13418.
  3. 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…
  4. Andrea Gemelli, Simone Marinai, Lorenzo Pisaneschi, and Francesco Santoni. 2024. Datasets and annotations for layout analysis of scientific articles. International Journal on Document Analysis and Recognition (IJDAR), 27(4):683–705. find
  5. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, and 1 others. 2025. Deepseek-R1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948.
  6. Caijun Jia, Nan Xu, Jingxuan Wei, Qingli Wang, Lei Wang, Bihui Yu, and Junnan Zhu. 2025. Chartrea-soner: Code-driven modality bridging for long-chain reasoning in chart question answering. arXiv preprint arXiv:2506.10116. — [arXiv:2506.10116](https://arxiv.org/abs/2506.10116) find
  7. Nan Jiang, Shanchao Liang, Chengxiao Wang, Jiannan Wang, and Lin Tan. 2025. Latte: Improving latex recognition for tables and formulae with iterative refinement. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 4030– 4038. find
  8. Yuxiang Lai, Jike Zhong, Ming Li, Shitian Zhao, and Xiaofeng Yang. 2025. Med-R1: Reinforcement learning for generalizable medical reasoning in vision-language models. arXiv preprint arXiv:2503.13939. — [arXiv:2503.13939](https://arxiv.org/abs/2503.13939) find
  9. Kenton Lee, Mandar Joshi, Iulia Raluca Turc, Hexiang Hu, Fangyu Liu, Julian Martin Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. 2023. Pix2struct: Screenshot parsing as pretraining for visual language understanding. In International Conference on Machine Learni…
  10. Jun Ling, Yao Qi, Tao Huang, Shibo Zhou, Yan-qin Huang, Jiang Yang, Ziqi Song, Ying Zhou, Yang Yang, Heng Tao Shen, and 1 others. 2025. Table2latex-rl: High-fidelity latex code generation from table images via reinforced multimodal language models. arXiv preprint arXiv:2509.17589. — [arXiv:2509.17589](https://arxiv.org/abs/2509.17589) find
  11. Aixin Liu, Aoxue Mei, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, and 1 others. 2025. Deepseek-v3. 2: Pushing the frontier of open large language models. arXiv preprint arXiv:2512.02556.
  12. OpenAI. 2025a. Gpt-4o: Gpt-4 with vision capabilities. https://openai.com/research/gpt-4o. ↗
  13. OpenAI. 2025b. Introducing gpt-5.2. https://openai.com/index/introducing-gpt-5-2. ↗
  14. Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, and 1 others. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730–27744.
  15. Dilruk Perera, Gousia Habib, Qianyi Xu, Daniel J Tan, Kai He, Erik Cambria, and Mengling Feng. 2025. Beyond prediction: Reinforcement learning as the defining leap in healthcare ai. arXiv preprint arXiv:2508.21101. — [arXiv:2508.21101](https://arxiv.org/abs/2508.21101) find
  16. Xiaoye Qu, Yafu Li, Zhaochen Su, Weigao Sun, Jianhao Yan, Dongrui Liu, Ganqu Cui, Daizong Liu, Shuxian Liang, Junxian He, and 1 others. 2025. A survey of efficient reasoning for large reasoning models: Language, multimodality, and beyond. arXiv preprint arXiv:2503.21614.
  17. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, and 1 others. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300.
  18. Lingxiao Tang, He Ye, Zhongxin Liu, Xiaoxue Ren, and Lingfeng Bao. 2025. Codereasoner: Enhancing the code reasoning ability with reinforcement learning. arXiv preprint arXiv:2507.17548. — [arXiv:2507.17548](https://arxiv.org/abs/2507.17548) find
  19. Renqiu Xia, Song Mao, Xiangchao Yan, Hongbin Zhou, Bo Zhang, Haoyang Peng, Jiahao Pi, Daocheng Fu, Wenjie Wu, Hancheng Ye, and 1 others. 2024. Docgenome: An open large-scale scientific document benchmark for training and testing multi-modal large language models. arXiv preprint arXiv:2406.11633. — [arXiv:2406.11633](https://arxiv.org/abs/2406.11633) find
  20. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388.
  21. Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Weinan Dai, Tiantian Fan, Gaohong Liu, Lingjun Liu, and 1 others. 2025a. DAPO: An open-source llm reinforcement learning system at scale. arXiv preprint arXiv:2503.14476.
  22. Tao Yu, Yi-Fan Zhang, Chaoyou Fu, Junkang Wu, Jinda Lu, Kun Wang, Xingyu Lu, Yunhang Shen, Guibin Zhang, Dingjie Song, and 1 others. 2025b. Aligning multimodal llm with human preference: A survey. arXiv preprint arXiv:2503.14504. — [arXiv:2503.14504](https://arxiv.org/abs/2503.14504) find
  23. Xu Zhong, Elaheh ShafieiBavani, and Antonio Jimeno Yepes. 2020. Image-based table recognition: data, model, and evaluation. In European conference on computer vision, pages 564–580. Springer. find
  24. Guanghao Zhou, Panjia Qiu, Cen Chen, Jie Wang, Zheming Yang, Jian Xu, and Minghui Qiu. 2025. Reinforced mllm: A survey on rl-based reasoning in multimodal large language models. arXiv preprint arXiv:2504.21277. — [arXiv:2504.21277](https://arxiv.org/abs/2504.21277) find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/yang-2026-cspo-alleviating-reward-ambiguity-structured/