Elamur Research⌕
☰
home / papers / 2026 / Accurate Legal Reasoning at Scale: Neuro-Symbolic Offloading and Structural Auditability for Robust Legal Adjudication

Accurate Legal Reasoning at Scale: Neuro-Symbolic Offloading and Structural Auditability for Robust Legal Adjudication

Paper · 2026 · Stanisław Sójka and Witold Kowalczyk

Summary

Large reasoning models (LRMs) can explain contract or statutory provisions that involve formulas, thresholds and dates. Using them to decide every individual transaction, however, is expensive and unreliable. Arithmetic slips, dropped conditions and different answers to identical inputs are unacceptable in billing or legal decisions. The authors (Delos AI) propose Amortized Intelligence. An LLM runs once, as a compiler, to convert a contract into DACL (Deterministic Autonomous Contract Language), a typed graph representation built from a small set of clause primitives. Each later event is resolved by executing this graph deterministically. A lightweight orchestrator (gpt-5-mini) only selects the relevant clauses and phrases the answer, and every decision comes with an inspectable execution trace. The authors tested the system on four proprietary industrial agreements against GPT-5.2, Claude 4.5 Sonnet and Gemini 3 Pro with reasoning enabled. The DACL Agent reaches about 99.5% accuracy, avoids the accuracy collapse LRMs show on complex clauses, and runs far cheaper and faster. This matters for high-volume, audit-sensitive automation where a human cannot review every output.

Key points

  • DACL represents contract logic as a DAG with four recursive primitives: Procedure, Logical Clause, Range Clause and Pricing Formula. Clause-level validity dates let amendments be recompiled locally.
  • The runtime agent is built on the OpenAI Agents SDK and has a single tool (evaluate_clauses_tool) that calls the deterministic engine.
  • On efficiency, the DACL Agent used 1.35M tokens versus 13.44M for GPT-5.2 Medium (9.9x fewer). On Logistics-MSA it took about 26.8 s per event versus roughly 164 s for the reasoning baselines. The paper reports cost savings of over 90% in the abstract and over 95% for high-volume tasks in its contributions list. GPT-5.2 without reasoning reached only 58% accuracy.
  • The system has been in production for more than 12 months, handling billing and accounts payable for 150+ agreements and about 1,000 billing events per month.
  • The authors note several limitations:
    • The primitives do not cover defeasible reasoning or open-textured standards.
    • The evaluation uses synthetic traffic and English contracts only.
    • Any compile-time misreading of a clause is repeated deterministically until it is fixed.

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 · 16,667 worksArtificial Intelligence · 16,194 worksSocial Sciences · 1,772 worksPolitical Science and International Relations · 182 works

🧭 Topics 2Research subjects and languages the work is about.Reasoning · 1,426 worksLegal AI · 134 works

Cite

Cite as Stanisław Sójka and Witold Kowalczyk. 2026. Accurate Legal Reasoning at Scale: Neuro-Symbolic Offloading and Structural Auditability for Robust Legal Adjudication. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-industry.102

@inproceedings{sojka2026accurate,
  title = {{Accurate Legal Reasoning at Scale: Neuro-Symbolic Offloading and Structural Auditability for Robust Legal Adjudication}},
  author = {Stanisław Sójka and Witold Kowalczyk},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-industry.102},
  url = {https://doi.org/10.18653/v1/2026.acl-industry.102}
}

References

42 references, 12 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Nikolaos Aletras, Dimitrios Tsarapatsanis, Daniel Preotiuc-Pietro, and Vasileios Lampos. 2016. Predicting judicial decisions of the european court of human rights: a natural language processing perspective. PeerJ Comput. Sci., 2:e93. find
  2. Anthropic. 2025. Claude sonnet 4.5 system card. Technical report, Anthropic. Introducing Extended Thinking capabilities. find
  3. Grigoris Antoniou and Antonis Bikakis. 2007. Dr-prolog: A system for defeasible reasoning with rules and ontologies on the semantic web. IEEE Transactions on Knowledge and Data Engineering, 19(2):233–245. ↗
  4. Andrew Blair-Stanek, Nils Holzenberger, and Benjamin Van Durme. 2023. Can gpt-3 perform statutory reasoning? In Proceedings of the Nineteenth International Conference on Artificial Intelligence and Law, ICAIL ’23, page 22–31, New York, NY, USA. Association for Computing Machinery.
  5. Ilias Chalkidis, Ion Androutsopoulos, and Nikolaos Aletras. 2019. Neural legal judgment prediction in English. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4317–4323, Florence, Italy. Association for Computational Linguistics.
  6. Ilias Chalkidis, Abhik Jana, Dirk Hartung, Michael Bommarito, Ion Androutsopoulos, Daniel Katz, and Nikolaos Aletras. 2022. LexGLUE: A benchmark dataset for legal language understanding in English. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4310–4330, Dublin, Ireland. Association for Computational Linguistics. ↗
  7. Lingjiao Chen, Matei A. Zaharia, and James Y. Zou. 2023. Frugalgpt: How to use large language models while reducing cost and improving performance. ArXiv, abs/2305.05176. ↗
  8. Wenhu Chen, Xueguang Ma, Xinyi Wang, and William W. Cohen. 2022. Program of thoughts prompting: Disentangling computation from reasoning for numerical reasoning tasks. Trans. Mach. Learn. ↗
  9. Jiaxi Cui, Zongjia Li, Yang Yan, Bohua Chen, and Li Yuan. 2023. Chatlaw: A multi-agent collaborative legal assistant with knowledge graph enhanced mixture-of-experts large language model. find
  10. Luc De Raedt, Angelika Kimmig, and Hannu Toivonen. 2007. Problog: A probabilistic prolog and its application in link discovery. In IJCAI’07: Proceedings of the 20th International Joint Conference on Artificial Intelligence, pages 2462–2467. find
  11. Moriya Dechtiar, Daniel Martin Katz, Sylvain Jaume, and Hongming Wang. 2025a. Llm as a judge for evaluating contract graphs: Multi-judge benchmarking and agentic uncertainty-aware refinement. Also available via SSRN: https://ssrn.com/abstract=5937996. ↗
  12. Moriya Dechtiar, Daniel Martin Katz, Mari Sundaresan, Sylvain Jaume, and Hongming Wang. 2025b. Graph-grpo-lex: Contract graph modeling and reinforcement learning with group relative policy optimization. arXiv preprint arXiv:2511.06618. ↗
  13. Google DeepMind. 2025. Gemini 3 pro model card. Technical report, Google. Version: gemini-3-pro-preview. find
  14. Yu Fan, Jingwei Ni, Jakob Merane, Yang Tian, Yoan Hermstrüwer, Yinya Huang, Mubashara Akhtar, Etienne Salimbeni, Florian Geering, Oliver Dreyer, Daniel Brunner, Markus Leippold, Mrinmaya Sachan, Alexander Stremitzer, Christoph Engel, Elliott Ash, and Joel Niklaus. 2026. Lexam: Benchmarking legal reasoning on 340 law exams. Preprint, arXiv:2505.12864. ↗
  15. Luyu Gao, Aman Madaan, Shuyan Zhou, Uri Alon, Pengfei Liu, Yiming Yang, Jamie Callan, and Graham Neubig. 2023. Pal: program-aided language models. In Proceedings of the 40th International Conference on Machine Learning, ICML’23. JMLR.org. ↗
  16. Saibo Geng, Martin Josifosky, Maxime Peyrard, and Robert West. 2023. Grammar-constrained decoding for structured nlp tasks without finetuning. In Conference on Empirical Methods in Natural Language Processing. find
  17. Neel Guha, Julian Nyarko, Daniel E. Ho, Christopher Ré, Adam Chilton, Aditya Narayana, Alex Chohlas-Wood, Austin M. K. Peters, Brandon Waldon, Daniel N. Rockmore, Diego A. Zambrano, Dmitry Talisman, Enam Hoque, Faiz Surani, Frank Fagan, Galit Sarfaty, Gregory M. Dickinson, Haggai Porat, Jason Hegland, and 21 others. 2023. Legalbench: A collaboratively built benchmark for measuring legal reasoning in large language models. ArXiv, abs/2308.11462. ↗
  18. Dan Hendrycks, Collin Burns, Anya Chen, and Spencer Ball. 2021. Cuad: An expert-annotated nlp dataset for legal contract review. Preprint, arXiv:2103.06268. ↗
  19. Abe Bohan Hou, Orion Weller, Guanghui Qin, Eugene Yang, Dawn Lawrie, Nils Holzenberger, Andrew Blair-Stanek, and Benjamin Van Durme. 2025. CLERC: A dataset for U. S. legal case retrieval and retrieval-augmented analysis generation. In Findings of the Association for Computational Linguistics: NAACL 2025, pages 7913–7928, Albuquerque, New Mexico. Association for Computational Linguistics. find
  20. Cheng-Yu Hsieh, Chun-Liang Li, Chih-kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alex Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. 2023. Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes. In Findings of the Association for Computational Linguistics: ACL 2023, pages 8003–8017, Toronto, Canada. Association for Computational Linguistics. ↗
  21. Quzhe Huang, Mingxu Tao, Zhenwei An, Chen Zhang, Cong Jiang, Zhibin Chen, Zirui Wu, and Yansong Feng. 2023. Lawyer llama technical report. CoRR, abs/2305.15062. ↗
  22. Manuj Kant, Manav Kant, Marzieh Nabi, Preston Carlson, and Megan Ma. 2024. Equitable access to justice: Logical llms show promise. arXiv preprint arXiv:2410.09904. ↗
  23. Manuj Kant, Sareh Nabi, Manav Kant, Roland Scharrer, Megan Ma, and Marzieh Nabi. 2025. Towards robust legal reasoning: Harnessing logical llms in law. Preprint, arXiv:2502.17638. ↗
  24. Daniel Martin Katz, Michael J Bommarito, and Josh Blackman. 2017. A general approach for predicting the behavior of the supreme court of the united states. PloS one, 12(4):e0174698. ↗
  25. Philipp Körner, Michael Leuschel, João Barbosa, Vítor Santos Costa, Verónica Dahl, Manuel V. Hermenegildo, Jose F. Morales, Jan Wielemaker, Daniel Diaz, Salvador Abreu, and Giovanni Ciatto. 2022. Fifty years of prolog and beyond. Theory and Practice of Logic Programming, 22(6):776–858. ↗
  26. Spyretta Leivaditi, Julien Rossi, and E. Kanoulas. 2020. A benchmark for lease contract review. ArXiv, abs/2010.10386. ↗
  27. Jason Morris. 2023. Building blawx. In 3rd International Workshop on Goal-Directed Execution of Answer Set Programs (GDE 2023). find
  28. OpenAI. 2025a. Gpt-5 mini: Lightweight reasoning for agentic workflows. Model version: gpt-5-mini-2025-12-11. find
  29. OpenAI. 2025b. Gpt-5.2 system card. Technical report, OpenAI. Model version: gpt-5.2-2025-12-11. find
  30. OpenAI. 2025c. Openai agents sdk: A framework for orchestrating intelligent workflows. https://github.com/openai/openai-agents-python. Software library. ↗
  31. Richard Shin, Christopher Lin, Sam Thomson, Charles Chen, Subhro Roy, Emmanouil Antonios Platanios, Adam Pauls, Dan Klein, Jason Eisner, and Benjamin Van Durme. 2021. Constrained language models yield few-shot semantic parsers. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7699–7715, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. ↗
  32. Miles Turpin, Julian Michael, Ethan Perez, and Samuel R. Bowman. 2023. Language models don’t always say what they think: unfaithful explanations in chain-of-thought prompting. In Proceedings of the 37th International Conference on Neural Information Processing Systems, NIPS ’23, Red Hook, NY, USA. Curran Associates Inc.
  33. Santosh T.y.s.s, Oana Ichim, and Matthias Grabmair. 2023. Zero-shot transfer of article-aware legal outcome classification for European court of human rights cases. In Findings of the Association for Computational Linguistics: EACL 2023, pages 605–617, Dubrovnik, Croatia. Association for Computational Linguistics. ↗
  34. Priyesh Vakharia, Abigail Kufeldt, Max Meyers, Ian Lane, and Leilani Gilpin. 2024. Proslm: A prolog synergized language model for explainable domain specific knowledge based question answering. arXiv preprint arXiv:2409.11589. ↗
  35. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2022. React: Synergizing reasoning and acting in language models. ArXiv, abs/2210.03629. ↗
  36. António Porto. 1984. Epilog: A language for extended programming in logic. In Implementations of Prolog, pages 268–278. find
  37. Kepu Zhang, Guofu Xie, Weijie Yu, Mingyue Xu, Xu Tang, Yaxin Li, and Jun Xu. 2025a. Legal Mathematical Reasoning with LLMs: Procedural Alignment through Two-Stage Reinforcement Learning. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 1586–1602, Singapore. Association for Computational Linguistics. find
  38. Albert Sadowski and Jarosław A. Chudziak. 2025a. Explainable rule application via structured prompting: A neural-symbolic approach. arXiv preprint arXiv:2506.16335. ↗
  39. Li Zhang, Matthias Grabmair, Morgan A. Gray, and Kevin Ashley. 2025b. Thinking longer, not always smarter: Evaluating llm capabilities in hierarchical legal reasoning. ArXiv, abs/2510.08710. ↗
  40. Albert Sadowski and Jarosław A. Chudziak. 2025b. On verifiable legal reasoning: A multi-agent framework with formalized knowledge representations. In Proceedings of the 34th ACM International Conference on Information and Knowledge Management (CIKM), pages —. Also available on arXiv:2509.00710. ↗
  41. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. Judging llm-as-a-judge with mt-bench and chatbot arena. In Proceedings of the 37th International Conference on Neural Information Processing Systems, NIPS ’23, Red Hook, NY, USA. Curran Associates Inc.
  42. Timo Schick, Jane Dwivedi-Yu, Roberto Dessí, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023. Toolformer: language models can teach themselves to use tools. In Proceedings of the 37th International Conference on Neural Information Processing Systems, NIPS ’23, Red Hook, NY, USA. Curran Associates Inc.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/sojka-2026-accurate-legal-reasoning-scale-neuro/