Elamur Research⌕
☰
home / papers / 2026 / NITI: Neural Plan Concretization for Incremental Execution, Bridging and Trigger Inference from Underspecified Human Policies

NITI: Neural Plan Concretization for Incremental Execution, Bridging and Trigger Inference from Underspecified Human Policies

Summary

The paper examines how to make large language models reliably execute abstract human policies, such as clinical guidelines. Such policies deliberately do not enumerate all preconditions and contingencies. The authors propose NITI, a neurosymbolic framework in which the LLM does not plan on its own but instead instantiates a given policy step by step at execution time. NITI has four mechanisms. Formal parameterized action schemas describe how the state changes. Trigger inference determines which stage of the policy corresponds to the current state, while short-circuiting skips subgoals that have already been achieved. A verifier provides feedback used to repair actions. Neural bridging returns the system to a point where the policy is applicable again. If the state is inconsistent, execution halts safely. The framework operates solely through in-context learning, without fine-tuning. It was evaluated on a new benchmark of 50 official WCA scrambles of the 2×2 cube and on 102 automated insulin delivery (AID) scenarios. Baseline methods solve none of the scrambles, whereas NITI with Claude Opus 4.5 solves 94% and nearly matches classical MPC in safety. The authors conclude that LLM planners lack execution structure rather than model capacity.

Key points

  • WCA 2×2 benchmark: solutions are typically longer than 10 moves, whereas in Sys2Bench they are mostly shorter than 4. With NITI, the share of solved scrambles is 94% for Claude Opus 4.5, 76% for GPT-5.2, and 36% for GPT-o4-mini. One-shot, CoT, and PoT yield 0%.
  • AID: the share of safe runs increased from 41.2% (one-shot) to 96.1% (NITI + Claude Opus 4.5), compared with 97% for MPC. Time in Range is 81% versus 78.2% for MPC. The share of hallucinated corrective actions decreased from 31.6% to 1.1%.
  • In the ablation, neural bridging proved the most important: without it, Opus's result drops from 94 to 34 points. Verifier-grounded repair, trigger inference, and short-circuiting follow in order of impact.
  • Overhead: a one-time learning phase requires 35 model calls, after which each task needs only a single call.
  • The code and benchmark are open: github.com/ImpactLabASU/ACLNITI.

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: Linguistics & languages

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.2007

@inproceedings{banerjee2026niti,
  title = {{NITI: Neural Plan Concretization for Incremental Execution, Bridging and Trigger Inference from Underspecified Human Policies}},
  author = {Banerjee, Ayan and Banerjee, Shomrik Barua and Gupta, Sandeep},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.2007},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.2007}
}

References

33 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Jonathan C. Balloch, Zhiyu Lin, Xiangyu Peng, Mustafa Hussain, Aarun Srinivas, Robert Wright, Julia M. Kim, and Mark O. Riedl. 2023. Neuro-symbolic world models for adapting to open world novelty. In Proceedings of the 2023 International Conference on Autonomous Agents and Multiagent Systems, AA-MAS… find
  2. Ayan Banerjee and Sandeep Gupta. 2024a. Emily: Extracting sparse model from implicit dynamics. In Proceedings of the 1st ECAI Workshop on "Machine Learning Meets Differential Equations: From Theory to Applications", volume 255 of Proceedings of Machine Learning Research, pages 1–11. PMLR. find
  3. Ayan Banerjee and Sandeep Gupta. 2025. Personalized open world plan generation for safety-critical human centered autonomous systems: A case study on artificial pancreas. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 22409–22422, Suzhou, China. Association for Compu… find
  4. Ayan Banerjee and Sandeep KS Gupta. 2024b. Recovering implicit physics model under real-world constraints. In European conference on Artificial Intelligence, Oral Presentation. arxiv:2412.02215. — [arXiv:2412.02215](https://arxiv.org/abs/2412.02215) find
  5. Ayan Banerjee, Aranyak Maity, Imane Lamrani, and Sandeep KS Gupta. 2025. Towards certified safe personalization in learning-enabled human-in-the-loop human-in-the-plant systems. ACM Journal on Emerging Technologies in Computing Systems, 22(1):1–27. find
  6. Richard N Bergman. 2021. Origins and history of the minimal model of glucose regulation. Frontiers in endocrinology, 11:583016. find
  7. Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. 2016. OpenAI gym. arXiv preprint arXiv:1606.01540. find
  8. Wenhu Chen, Xueguang Ma, Xinyi Wang, and William W. Cohen. 2023. Program of thoughts prompting: Disentangling computation from reasoning for numerical reasoning tasks. Preprint, arXiv:2211.12588. find our card
  9. Bruce D. D’Ambrosio. 1987. Combining symbolic and numeric approaches to uncertainty management. In Proceedings of the Third Conference on Uncertainty in Artificial Intelligence, UAI’87, page 386–393, Arlington, Virginia, USA. AUAI Press. find
  10. Venkata Naren Devarakonda, Raktim Gautam Goswami, Ali Umut Kaypak, Naman Patel, Rooholla Khorram-bakht, Prashanth Krishnamurthy, and Farshad Khor-rami. 2024. Orionnav: Online planning for robot autonomy with context-aware llm and open-vocabulary semantic scene graphs. Preprint, arXiv:2410.06239. — [arXiv:2410.06239](https://arxiv.org/abs/2410.06239) find
  11. Subhabrata Dutta, Timo Kaufmann, Goran Glavaš, Ivan Habernal, Kristian Kersting, Frauke Kreuter, Mira Mezini, Iryna Gurevych, Eyke Hüllermeier, and Hinrich Schuetze. 2025. Problem solving through human–ai preference-based cooperation. Computational Linguistics, pages 1–36. find
  12. Kazuki Hori, Kanata Suzuki, and Tetsuya Ogata. 2024. Interactively robot action planning with uncertainty analysis and active questioning by large language model. In 2024 IEEE/SICE International Symposium on System Integration (SII), pages 85–91. find
  13. Michael Katz, Harsha Kokel, Christian Muise, Shirin Sohrabi, and Sarath Sreedharan. 2025. Make planning research rigorous again! Preprint, arXiv:2505.21674. — [arXiv:2505.21674](https://arxiv.org/abs/2505.21674) find
  14. Ruoyu Li and James T. Allison. 2017. Nonlinear optimization for planning problems with complex constraints and objectives. IEEE Transactions on Automation Science and Engineering, 14(3):1419– 1432. find
  15. Kaiqu Liang, Zixu Zhang, and Jaime Fernández Fisac. 2024. Introspective planning: Aligning robots’ uncertainty with inherent task ambiguity. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. find
  16. Xiaotian Liu, Ali Pesaranghader, Hanze Li, Punyaphat Sukcharoenchaikul, Jaehong Kim, Tanmana Sadhu, Hyejeong Jeon, and Scott Sanner. 2025. Open-world planning via lifted regression with LLM-inferred affordances for embodied agents. In Proceedings of the 63rd Annual Meeting of the Association for Com… find
  17. Runyu Ma, Jelle Luijkx, Zlatan Ajanović, and Jens Kober. 2024. ExploRLLM: Guiding exploration in reinforcement learning with large language models. In RSS 2024 Workshop: Data Generation for Robotics. find
  18. Chiara Dalla Man, Francesco Micheletto, Dayu Lv, Marc Breton, Boris Kovatchev, and Claudio Cobelli. 2014. The uva/padova type 1 diabetes simulator: new features. Journal of diabetes science and technology, 8(1):26–34. find
  19. Jiayuan Mao, Tomás Lozano-Pérez, Joshua B. Tenenbaum, and Leslie Pack Kaelbling. 2023. What planning problems can a relational neural network solve? In Proceedings of the 37th International Conference on Neural Information Processing Systems, NIPS ’23, Red Hook, NY, USA. Curran Associates Inc. find
  20. Riddell MC, Gallen IW, Smart CE, Taplin CE, Adolfsson P, Lumb AN, Kowalski A, Rabasa-Lhoret R, McCrimmon RJ, Hume C, Annan F, Fournier PA, Graham C, Bode B, Galassetti P, Jones TW, Millán IS, Heise T, Peters AL, and 2 others. 2017. Exercise management in type 1 diabetes: a consensus statement. Lance… find
  21. Epub 2017 Jan 24. Erratum in: Lancet Diabetes Endocrinol. 2017 May;5(5):e3. doi: 10.1016/S2213-8587(17)30086-4. — [doi:10.1016/s2213-8587(17)30086-4](https://doi.org/10.1016/s2213-8587%2817%2930086-4) find
  22. Héctor Palacios and Hector Geffner. 2016. Learning action models for re-planning. In Proceedings of the 25th International Joint Conference on Artificial Intelligence (IJCAI), pages 1670–1676. find
  23. Shubham Parashar, Blake Olson, Sambhav Khurana, Eric Li, Hongyi Ling, James Caverlee, and Shui-wang Ji. 2025. Inference-time computations for llm reasoning and planning: A benchmark and insights. Preprint, arXiv:2502.12521. — [arXiv:2502.12521](https://arxiv.org/abs/2502.12521) find
  24. Cassandra Parent, Elodie Lespagnol, Serge Berthoin, Sémah Tagougui, Joris Heyman, Chantal Stuckens, Iva Gueorguieva, Costantino Balestra, Cajsa Tonoli, Bérengère Kozon, and 1 others. 2023. Barriers to physical activity in children and adults living with type 1 diabetes: A complex link with real-life… find
  25. Deepak Pathak, Pulkit Agrawal, Alexei A. Efros, and Trevor Darrell. 2017. Curiosity-driven exploration by self-supervised prediction. In Proceedings of the 34th International Conference on Machine Learning - Volume 70, ICML’17, page 2778–2787. JMLR.org. find
  26. Navid Resalat, Wade Hilts, Joseph El Youssef, Nic-hole Tyler, Jessica R Castle, and Peter G Jacobs. 2019. Adaptive control of an artificial pancreas using model identification, adaptive postprandial insulin delivery, and heart rate and accelerometry as control inputs. Journal of Diabetes Science and… find
  27. Michael C Riddell, Ian W Gallen, Carmel E Smart, Craig E Taplin, Peter Adolfsson, Alistair N Lumb, Aaron Kowalski, Remi Rabasa-Lhoret, Rory J Mc-Crimmon, Carin Hume, and 1 others. 2017. Exercise management in type 1 diabetes: a consensus statement. The lancet Diabetes & endocrinology, 5(5):377–390. find
  28. Juan A. Rodriguez-Aguilar, Ulle Endriss, Sarvapali D. Ramchurn, and Michael Luck. 2010. Efficient genetic algorithms for multi-agent plan coordination. Journal of Artificial Intelligence Research, 39:59– 101. find
  29. Jun Wang, Guocheng He, and Yiannis Kantaros. 2024. Safe task planning for language-instructed multi-robot systems using conformal prediction. CoRR, abs/2402.15368. — [arXiv:2402.15368](https://arxiv.org/abs/2402.15368) find
  30. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models. In Proceedings of the 36th International Conference on Neural Information Processing Systems, NIPS ’22, R… find our card
  31. Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L Griffiths, Yuan Cao, and Karthik Narasimhan. 2024. Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems, 36. find
  32. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629. find our card
  33. Yanfang Zhou, Yuntao Liu, Xiaodong Li, Yongqiang Zhao, Xintong Wang, Jinlong Tian, Zhenyu Li, and Xinhai Xu. 2025. Metagent-P: A neuro-symbolic planning agent with metacognition for open worlds. In Findings of the Association for Computational Linguistics: ACL 2025, pages 22747–22764, Vienna, Austri… find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/papers/banerjee-2026-niti-neural-plan-concretization-incremental/