Elamur Research⌕
☰
home / papers / 2026 / Reference-Free Schema Generation for Literature Review Tables via Multi-Faceted Rewards

Reference-Free Schema Generation for Literature Review Tables via Multi-Faceted Rewards

Paper · 2026 · Sinjoy Saha, Suman Saha, Mahfuza Farooque et al.

Summary

Building literature review tables involves two steps: choosing comparison columns (the schema) and filling in their values. The first step has no clean match to a standard NLP task, and gold schemas for supervised training are scarce. The authors treat schema induction as reinforcement learning that needs no reference schemas. Their reward combines four uniformly weighted signals:

  • NLI-based entailment of each attribute against every paper,
  • a capped count of attributes to reward coverage,
  • embedding dissimilarity between attributes to discourage redundancy,
  • a graded check of JSON structure.

They train Qwen3-0.6B with GRPO on 1,730 arxivDIGESTables instances, using only titles and abstracts, and evaluate on the 100-table arxivDIGESTables-Clean subset. The tuned model beats its prompted base on most intrinsic signals, on BERTScore-based alignment and on LLM-judge ratings. The 3B SFT baselines remain ahead on reference matching and corpus grounding. The RL model's attribute diversity and judge scores are close to the SFT models', although its clarity rating lags noticeably. This suggests reward design can partly replace costly schema annotation.

Key points

  • Reward components:
    • Entailment uses the cross-encoder nli-MiniLM2-L6-H768, scoring each attribute (name plus short description) against each document.
    • Coverage is min(n/N_max, 1).
    • Dissimilarity is mean pairwise cosine dissimilarity computed with all-MiniLM-L6-v2.
    • Format is scored in sequential steps: 0.3 for valid JSON, 0.2 for a non-empty dict, 0.3 for required fields present, 0.2 for valid field types.
  • Gains over the prompted Qwen3-0.6B baseline:
    • Format rose from 0.86 to 0.91, coverage from 0.49 to 0.51 and dissimilarity from 0.250 to 0.259.
    • Entailment slightly dropped, from 0.047 to 0.042.
    • BERTScore AUC F1 increased from 0.156 to 0.175.
  • LLM-judge results (GPT-5-mini) for the RL model: relevance 0.77, usefulness 0.66, clarity 0.60. The strongest SFT baseline (Qwen 2.5 3B) scores 0.81, 0.71 and 0.78, with an F1 of 0.243.
  • SFT baselines are Llama 3.2 3B and Qwen 2.5 3B trained on gold tables (Padmakumar et al., 2025). Both have about 5× the parameters of the RL model.
  • Observed errors and limitations:
    • Some outputs contain unescaped quotes that break the JSON.
    • Some output formats are vague, such as "numbers or strings", yet still pass the format reward.
    • Other limitations include fixed reward weights, a risk of reward hacking, and evaluation on only 100 tables without human judgments.

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

🎓Pennsylvania State UniversityState College · United States56 works · 5,360 citations

 Fields, topics and concepts

📚 Shelves 6The 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 worksSocial Sciences · 1,772 worksGeneral Social Sciences · 227 works

Cite

Cite as Sinjoy Saha, Suman Saha, Mahfuza Farooque and Wenpeng Yin. 2026. Reference-Free Schema Generation for Literature Review Tables via Multi-Faceted Rewards. 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.111

@inproceedings{saha2026reference,
  title = {{Reference-Free Schema Generation for Literature Review Tables via Multi-Faceted Rewards}},
  author = {Sinjoy Saha, Suman Saha, Mahfuza Farooque and Wenpeng Yin},
  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.111},
  url = {https://doi.org/10.18653/v1/2026.acl-srw.111}
}

References

21 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Shubham Agarwal and 1 others. 2024. LitLLM: A toolkit for scientific literature review. In Proceedings of the 2024 Annual Conference of the North American Chapter of the Association for Computational Linguistics. Association for Computational Linguistics. find
  2. Anirudh Ajith, Mengzhou Xia, Alexis Chevalier, Tanya Goyal, Danqi Chen, and Tianyu Gao. 2024. Lit-Search: A retrieval benchmark for scientific literature search. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 15068–15083, Miami, Florida, USA. Associa… find
  3. Hosein Azarbonyad, Zi Long Zhu, Georgios Cheirmpos, Zubair Afzal, Vikrant Yadav, and Georgios Tsatsaronis. 2025. Question-answer extraction from scientific articles using knowledge graphs and large language models. In Proceedings of the 48th International ACM SIGIR Conference on Research and Develop… find
  4. Fan Bai, Junmo Kang, Gabriel Stanovsky, Dayne Freitag, Mark Dredze, and Alan Ritter. 2024. Schema-driven information extraction from heterogeneous tables. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 10252– 10273, Miami, Florida, USA. Association for Computational… 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 2015 Conference on Empirical Methods in Natural Language Processing, pages 632–642, Lisbon, Portugal. Association for Computati…
  6. Laura Chiticariu, Marina Danilevsky, Yunyao Li, Frederick Reiss, and Huaiyu Zhu. 2018. SystemT: Declarative text understanding for enterprise. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 3 (… find
  7. Raymond Fok, Joseph Chee Chang, Marissa Radensky, Pao Siangliulue, Jonathan Bragg, Amy X. Zhang, and Daniel S. Weld. 2025. Facets, taxonomies, and syntheses: Navigating structured representations in llm-assisted literature review. Preprint, arXiv:2504.18496. — [arXiv:2504.18496](https://arxiv.org/abs/2504.18496) find
  8. Hayato Hashimoto, Kazutoshi Shinoda, Hikaru Yokono, and Akiko Aizawa. 2017. Automatic generation of review matrices as multi-document summarization of scientific papers. In BIRNDL@SIGIR. find
  9. Omri Koshorek, Niv Granot, Aviv Alloni, Shahar Admati, Roee Hendel, Ido Weiss, Alan Arazi, Shay-Nitzan Cohen, and Yonatan Belinkov. 2025. Structured rag for answering aggregative questions. Preprint, arXiv:2511.08505. — [arXiv:2511.08505](https://arxiv.org/abs/2511.08505) find
  10. Yixin Liu, Yonghui Wu, Denghui Zhang, and Lichao Sun. 2025. Agentic autosurvey: Let llms survey llms. Preprint, arXiv:2509.18661. — [arXiv:2509.18661](https://arxiv.org/abs/2509.18661) find
  11. Benjamin Newman, Yoonjoo Lee, Aakanksha Naik, Pao Siangliulue, Raymond Fok, Juho Kim, Daniel S Weld, Joseph Chee Chang, and Kyle Lo. 2024. ArxivDI-GESTables: Synthesizing scientific literature into tables using language models. In Proceedings of the 2024 Conference on Empirical Methods in Natural La… find
  12. 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.
  13. Vishakh Padmakumar, Joseph Chee Chang, Kyle Lo, Doug Downey, and Aakanksha Naik. 2025. Intent-aware schema generation and refinement for literature review tables. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 23450–23472, Suzhou, China. Association for Computational… find
  14. Daniel M. Russell, Mark J. Stefik, Peter Pirolli, and Stuart K. Card. 1993. The cost structure of sensemaking. In Proceedings of the INTERACT ’93 and CHI ’93 Conference on Human Factors in Computing Systems, CHI ’93, page 269–276, New York, NY, USA. Association for Computing Machinery. find
  15. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Mingchuan Zhang, Y.K. Li, Y. Wu, and Daya Guo. 2024. DeepSeekMath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300.
  16. Aaditya Singh, Adam Fry, Adam Perelman, Adam Tart, Adi Ganesh, Ahmed El-Kishky, Aidan McLaughlin, Aiden Low, AJ Ostrow, Akhila Ananthram, Akshay Nathan, Alan Luo, Alec Helyar, Aleksander Madry, Aleksandr Efremov, Aleksandra Spyra, Alex Baker-Whitcomb, Alex Beutel, Alex Karpenko, and 465 others. 2025… — OpenAI GPT-5 System Card
  17. Amanpreet Singh, Joseph Chee Chang, Chloe Anastasiades, Dany Haddad, Aakanksha Naik, Amber Tanaka, Angele Zamarron, Cecile Nguyen, Jena D. Hwang, Jason Dunkleberger, Matt Latzke, Smita Rao, Jaron Lochner, Rob Evans, Rodney Kinney, Daniel S. Weld, Doug Downey, and Sergey Feldman. 2025b. Ai2 scholar q… find
  18. Liangtai Sun, Yang Han, Zihan Zhao, Da Ma, Zhennan Shen, Baocai Chen, Lu Chen, and Kai Yu. 2024. Scieval: A multi-level large language model evaluation benchmark for scientific research. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 19053–19061. find
  19. Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, and 41 others. 2025. Qwen3 technical report. Preprint, arXiv:2505.09388.
  20. Shuo Zhang and Krisztian Balog. 2018. On-the-fly table generation. In The 41st International ACM SIGIR Conference on Research amp; Development in Information Retrieval, SIGIR ’18, page 595–604. ACM. find
  21. Kun Zhu, Xiaocheng Feng, Xiachong Feng, Yingsheng Wu, and Bing Qin. 2023. Hierarchical catalogue generation for literature review: A benchmark. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 6790–6804, Singapore. Association for Computational Linguistics. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/saha-2026-reference-free-schema-generation-literature/