---
title: "Action Boundary Blindness — paper by Wang et al. (2026)"
description: "The paper examines a failure pattern in LLM agents that the authors call Action Boundary Blindness. 29 references."
authors: ["Zhangyi Wang", "Bingnan Yu", "Jiexiang Xu", "Zongze Li"]
year: 2026
venue: "Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)"
source: "https://aclanthology.org/2026.acl-long.1711/"
doi: "10.18653/v1/2026.acl-long.1711"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/wang-2026-action-boundary-blindness-when-llm/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/wang-2026-action-boundary-blindness-when-llm/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/21/) / Action Boundary Blindness: When LLM Agents Cannot Tell Where One Action Ends and Another Begins

# Action Boundary Blindness: When LLM Agents Cannot Tell Where One Action Ends and Another Begins

**Authors:** [Zhangyi Wang](https://md.research.elamur.ai/free/en/authors/zhangyi-wang/), [Bingnan Yu](https://md.research.elamur.ai/free/en/authors/bingnan-yu/), [Jiexiang Xu](https://md.research.elamur.ai/free/en/authors/jiexiang-xu/) and [Zongze Li](https://md.research.elamur.ai/free/en/authors/zongze-li/) · **Year:** 2026 · **Venue:** [Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)](https://md.research.elamur.ai/free/en/venues/acl/) ([ACL 2026](https://md.research.elamur.ai/free/en/venues/acl/2026/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2026.acl-long.1711/) · [PDF](https://aclanthology.org/2026.acl-long.1711.pdf) · [DOI 10.18653/v1/2026.acl-long.1711](https://doi.org/10.18653/v1/2026.acl-long.1711) · [OpenAlex](https://openalex.org/W7166884145)

**Citations:** 0 citations (OpenAlex, 2026-10-14)

## Summary
The paper examines a failure pattern in LLM agents that the authors call Action Boundary Blindness. An agent may choose the correct type of action but still do too little (skipping required steps), do too much (touching things outside the task), or split the work at the wrong granularity. Building on cognitive science's Event Segmentation Theory, the authors lay out a taxonomy of three violation categories, namely boundary ambiguity, scope creep, and granularity confusion. They also propose four metrics that need no human annotation: ABS, GAR, SVR, and BASR. They test ReAct-style agents built on seven LLMs over 1,655 tasks from six benchmarks, including τ-bench, WebArena, ALFWorld, TheAgentCompany, and OSWorld. Every model scores poorly on boundary quality, and boundary errors are the most common single cause of failure. A cheap intervention, Explicit Boundary Prompting (EBP), improves results for all models. This suggests the models already have some boundary awareness that default prompting does not draw out. The finding matters for enterprise and regulated deployments, where following the prescribed procedure is a safety requirement rather than a question of efficiency.

## Key points
- The top model reaches only 0.424 ABS, and the average across the seven LLMs is 0.393. Under-action is the most frequent violation type, at 48.4%.
- A multi-label attribution scheme (κ = 0.78) names boundary blindness as the primary failure cause in 37.2% of failed runs. It is the sole cause in 25.8% (351 of 1,361) and is involved in some way in 55.9%.
- BASR is about 4 points below the standard success rate. The gap reveals "lucky successes," where the outcome is correct but the process is flawed.
- EBP requires no training. It raises ABS by 0.08–0.13, raises success rate by 9.2%, and cuts violations by 42–47%.
- The metrics were checked with state-based cross-validation and a human audit of 150 samples (78.0% precision, estimated false positive rate of about 22%). Model rankings stayed the same (Spearman ρ = 1.0), and the LLM-as-Judge agreed with humans 93.8% of the time. All benchmarks are English-only.

*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

- [Columbia University](https://md.research.elamur.ai/free/en/universities/columbia-university/) — University, New York, United States · 241 works · 54,053 citations
- [Nanyang Technological University](https://md.research.elamur.ai/free/en/universities/nanyang-technological-university/) — University, Singapore, Singapore · 156 works · 25,644 citations

📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works

## Cite

Cite as: Zhangyi Wang, Bingnan Yu, Jiexiang Xu and Zongze Li. 2026. Action Boundary Blindness: When LLM Agents Cannot Tell Where One Action Ends and Another Begins. 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.

```bibtex
@inproceedings{wang2026action,
  title = {{Action Boundary Blindness: When LLM Agents Cannot Tell Where One Action Ends and Another Begins}},
  author = {Zhangyi Wang, Bingnan Yu, Jiexiang Xu and Zongze Li},
  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.1711},
  url = {https://doi.org/10.18653/v1/2026.acl-long.1711}
}
```

- [← previous ACR: Adaptive Context Refactoring via Context Refactoring Operators for Multi-Turn Dialogue](https://md.research.elamur.ai/free/en/papers/shen-2026-acr-adaptive-context-refactoring-context/)
- [next → Actionable Interpretability for Churn Classification: A Text Bottleneck Model Case Study at a Major Telecom Provider](https://md.research.elamur.ai/free/en/papers/sauter-2026-actionable-interpretability-churn-classification-text/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/21/)

## References

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

1. Matthew M Botvinick. 2008. Hierarchical models of behavior and prefrontal function. Trends in cognitive sciences, 12(5):201–208. [find](https://scholar.google.com/scholar?q=Matthew+M+Botvinick%2E+2008%2E+Hierarchical+models+of+behavior+and+prefrontal+function%2E+Trends+in+cognitive+sciences%2C+12%285%29%3A201%E2%80%93208%2E)
2. Mert Cemri, Melissa Z Pan, Shuyi Yang, Lakshya A Agrawal, Bhavya Chopra, Rishabh Tiwari, Kurt Keutzer, Aditya Parameswaran, Dan Klein, Kannan Ramchandran, and 1 others. 2025. Why do multi-agent llm systems fail? arXiv preprint arXiv:2503.13657. [find](https://scholar.google.com/scholar?q=Mert+Cemri%2C+Melissa+Z+Pan%2C+Shuyi+Yang%2C+Lakshya+A+Agrawal%2C+Bhavya+Chopra%2C+Rishabh+Tiwari%2C+Kurt+Keutzer%2C+Aditya+Parameswaran%2C+Dan+Klein%2C+Kannan+Ramchandran%2C+and+1+others%2E+2025%2E+Why+do+multi%2Dagent+llm+sy)
3. Darshan Deshpande, Varun Gangal, Hersh Mehta, Jitin Krishnan, Anand Kannappan, and Rebecca Qian. 2025. Trail: Trace reasoning and agentic issue localization. arXiv preprint arXiv:2505.08638. [find](https://scholar.google.com/scholar?q=Darshan+Deshpande%2C+Varun+Gangal%2C+Hersh+Mehta%2C+Jitin+Krishnan%2C+Anand+Kannappan%2C+and+Rebecca+Qian%2E+2025%2E+Trail%3A+Trace+reasoning+and+agentic+issue+localization%2E+arXiv+preprint+arXiv%3A2505%2E08638%2E)
4. Pengfei He, Zhenwei Dai, Bing He, Hui Liu, Xianfeng Tang, Hanqing Lu, Juanhui Li, Jiayuan Ding, Subhabrata Mukherjee, Suhang Wang, and 1 others. 2025. Traject-bench: A trajectory-aware benchmark for evaluating agentic tool use. arXiv preprint arXiv:2510.04550. — [arXiv:2510.04550](https://arxiv.org/abs/2510.04550) [find](https://scholar.google.com/scholar?q=Pengfei+He%2C+Zhenwei+Dai%2C+Bing+He%2C+Hui+Liu%2C+Xianfeng+Tang%2C+Hanqing+Lu%2C+Juanhui+Li%2C+Jiayuan+Ding%2C+Subhabrata+Mukherjee%2C+Suhang+Wang%2C+and+1+others%2E+2025%2E+Traject%2Dbench%3A+A+trajectory%2Daware+benchmark+for+e)
5. Panagiotis Michelakis, Yiannis Hadjiyiannis, and Dimitrios Stamoulis. 2025. Core: Full-path evaluation of llm agents beyond final state. arXiv preprint arXiv:2509.20998. — [arXiv:2509.20998](https://arxiv.org/abs/2509.20998) [find](https://scholar.google.com/scholar?q=Panagiotis+Michelakis%2C+Yiannis+Hadjiyiannis%2C+and+Dimitrios+Stamoulis%2E+2025%2E+Core%3A+Full%2Dpath+evaluation+of+llm+agents+beyond+final+state%2E+arXiv+preprint+arXiv%3A2509%2E20998%2E+%E2%80%94+%5BarXiv%3A2509%2E20998%5D%28https%3A%2F%2Fa)
6. Sebastian Michelmann, Manoj Kumar, Kenneth A Norman, and Mariya Toneva. 2025. Large language models can segment narrative events similarly to humans. Behavior Research Methods, 57(1):39. [find](https://scholar.google.com/scholar?q=Sebastian+Michelmann%2C+Manoj+Kumar%2C+Kenneth+A+Norman%2C+and+Mariya+Toneva%2E+2025%2E+Large+language+models+can+segment+narrative+events+similarly+to+humans%2E+Behavior+Research+Methods%2C+57%281%29%3A39%2E)
7. Saul B Needleman and Christian D Wunsch. 1970. A general method applicable to the search for similarities in the amino acid sequence of two proteins. Journal of molecular biology, 48(3):443–453. [find](https://scholar.google.com/scholar?q=Saul+B+Needleman+and+Christian+D+Wunsch%2E+1970%2E+A+general+method+applicable+to+the+search+for+similarities+in+the+amino+acid+sequence+of+two+proteins%2E+Journal+of+molecular+biology%2C+48%283%29%3A443%E2%80%93453%2E)
8. Juhyun Oh, Eunsu Kim, Inha Cha, and Alice Oh. 2024. The generative ai paradox in evaluation:“what it can solve, it may not evaluate”. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics: Student Research Workshop, pages 248–257. [find](https://scholar.google.com/scholar?q=Juhyun+Oh%2C+Eunsu+Kim%2C+Inha+Cha%2C+and+Alice+Oh%2E+2024%2E+The+generative+ai+paradox+in+evaluation%3A%E2%80%9Cwhat+it+can+solve%2C+it+may+not+evaluate%E2%80%9D%2E+In+Proceedings+of+the+18th+Conference+of+the+European+Chapter+of+t)
9. Archiki Prasad, Alexander Koller, Mareike Hartmann, Peter Clark, Ashish Sabharwal, Mohit Bansal, and Tushar Khot. 2024. [Adapt: As-needed decomposition and planning with language models](https://md.research.elamur.ai/free/en/papers/prasad-2024-adapt-as-needed-decomposition-planning/). In Findings of the Association for Computational Linguistics: NAACL 2024, pages 4226–4252.
10. Felix Hofstätter, Teun Van Der Weij, Jayden Teoh, Rada Djoneva, Henning Bartsch, and Francis Rhys Ward. 2025. The elicitation game: Evaluating capability elicitation techniques. arXiv preprint arXiv:2502.02180. — [arXiv:2502.02180](https://arxiv.org/abs/2502.02180) [find](https://scholar.google.com/scholar?q=Felix+Hofst%C3%A4tter%2C+Teun+Van+Der+Weij%2C+Jayden+Teoh%2C+Rada+Djoneva%2C+Henning+Bartsch%2C+and+Francis+Rhys+Ward%2E+2025%2E+The+elicitation+game%3A+Evaluating+capability+elicitation+techniques%2E+arXiv+preprint+arXiv%3A2)
11. Yaoyao Qian, Yuanli Wang, Jinda Zhang, Yun Zong, Meixu Chen, Hanhan Zhou, Jindan Huang, Yifan Zeng, Xinyu Hu, Chan Hee Song, and 1 others. 2025. Webgrapheval: Multi-turn trajectory evaluation for web agents using graph representation. arXiv preprint arXiv:2510.19205. — [arXiv:2510.19205](https://arxiv.org/abs/2510.19205) [find](https://scholar.google.com/scholar?q=Yaoyao+Qian%2C+Yuanli+Wang%2C+Jinda+Zhang%2C+Yun+Zong%2C+Meixu+Chen%2C+Hanhan+Zhou%2C+Jindan+Huang%2C+Yifan+Zeng%2C+Xinyu+Hu%2C+Chan+Hee+Song%2C+and+1+others%2E+2025%2E+Webgrapheval%3A+Multi%2Dturn+trajectory+evaluation+for+web+)
12. Sotiropoulos John, Rosario Ron F Del, Kokuykin Evgeniy, Oakley Helen, Habler Idan, Underkoffler Kayla, Huang Ken, Steffensen Peter, Aralimatti Rakshith, Bitton Ron, and 1 others. 2025. Owasp top 10 for llm apps & gen ai agentic security initiative. Ph.D. thesis, OWASP. [find](https://scholar.google.com/scholar?q=Sotiropoulos+John%2C+Rosario+Ron+F+Del%2C+Kokuykin+Evgeniy%2C+Oakley+Helen%2C+Habler+Idan%2C+Underkoffler+Kayla%2C+Huang+Ken%2C+Steffensen+Peter%2C+Aralimatti+Rakshith%2C+Bitton+Ron%2C+and+1+others%2E+2025%2E+Owasp+top+10+fo)
13. Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023. [Toolformer: Language models can teach themselves to use tools](https://md.research.elamur.ai/free/en/papers/schick-2023-toolformer-language-models-can-teach/). Advances in neural information processing systems, 36:68539–68551.
14. Manoj Kumar, Ariel Goldstein, Sebastian Michelmann, Jeffrey M Zacks, Uri Hasson, and Kenneth A Norman. 2023. Bayesian surprise predicts human event segmentation in story listening. Cognitive science, 47(10):e13343. [find](https://scholar.google.com/scholar?q=Manoj+Kumar%2C+Ariel+Goldstein%2C+Sebastian+Michelmann%2C+Jeffrey+M+Zacks%2C+Uri+Hasson%2C+and+Kenneth+A+Norman%2E+2023%2E+Bayesian+surprise+predicts+human+event+segmentation+in+story+listening%2E+Cognitive+science%2C+)
15. Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. [Reflexion: Language agents with verbal reinforcement learning](https://md.research.elamur.ai/free/en/papers/shinn-2023-reflexion-language-agents-verbal-reinforcement/). Advances in neural information processing systems, 36:8634–8652.
16. Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2023. [Let’s verify step by step](https://md.research.elamur.ai/free/en/papers/lightman-2023-let-s-verify-step-step/). In The twelfth international conference on learning representations.
17. Mohit Shridhar, Xingdi Yuan, Marc-Alexandre Côté, Yonatan Bisk, Adam Trischler, and Matthew Hausknecht. 2020. [Alfworld: Aligning text and embodied environments for interactive learning](https://md.research.elamur.ai/free/en/papers/shridhar-2020-alfworld-aligning-text-embodied-environments/). arXiv preprint arXiv:2010.03768.
18. Lei Wang, Wanyu Xu, Yihuai Lan, Zhiqiang Hu, Yunshi Lan, Roy Ka-Wei Lee, and Ee-Peng Lim. 2023. Plan-and-solve prompting: Improving zero-shot chain-of-thought reasoning by large language models. In Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: lon… [find](https://scholar.google.com/scholar?q=Lei+Wang%2C+Wanyu+Xu%2C+Yihuai+Lan%2C+Zhiqiang+Hu%2C+Yunshi+Lan%2C+Roy+Ka%2DWei+Lee%2C+and+Ee%2DPeng+Lim%2E+2023%2E+Plan%2Dand%2Dsolve+prompting%3A+Improving+zero%2Dshot+chain%2Dof%2Dthought+reasoning+by+large+language+models%2E+In+Pr)
19. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 2022. [Chain-of-thought prompting elicits reasoning in large language models](https://md.research.elamur.ai/free/en/papers/wei-2022-chain-thought-prompting-elicits-reasoning/). Advances in neural information processing systems, 35:24824– 24837.
20. Daniel M Wolpert, Zoubin Ghahramani, and Michael I Jordan. 1995. An internal model for sensorimotor integration. Science, 269(5232):1880–1882. [find](https://scholar.google.com/scholar?q=Daniel+M+Wolpert%2C+Zoubin+Ghahramani%2C+and+Michael+I+Jordan%2E+1995%2E+An+internal+model+for+sensorimotor+integration%2E+Science%2C+269%285232%29%3A1880%E2%80%931882%2E)
21. Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh J Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, and 1 others. 2024. [Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments](https://md.research.elamur.ai/free/en/papers/xie-2024-osworld-benchmarking-multimodal-agents-open/). Advances in Neural Information Processing Systems, 37…
22. Frank F Xu, Yufan Song, Boxuan Li, Yuxuan Tang, Kritanjali Jain, Mengxue Bao, Zora Z Wang, Xuhui Zhou, Zhitong Guo, Murong Cao, and 1 others. 2024. Theagentcompany: benchmarking llm agents on consequential real world tasks. arXiv preprint arXiv:2412.14161. [find](https://scholar.google.com/scholar?q=Frank+F+Xu%2C+Yufan+Song%2C+Boxuan+Li%2C+Yuxuan+Tang%2C+Kritanjali+Jain%2C+Mengxue+Bao%2C+Zora+Z+Wang%2C+Xuhui+Zhou%2C+Zhitong+Guo%2C+Murong+Cao%2C+and+1+others%2E+2024%2E+Theagentcompany%3A+benchmarking+llm+agents+on+conseque)
23. Shunyu Yao, Noah Shinn, Pedram Razavi, and Karthik Narasimhan. 2024. [τ -bench: A benchmark for tool-agent-user interaction in real-world domains](https://md.research.elamur.ai/free/en/papers/yao-2024-tau-bench-benchmark-tool/). arXiv preprint arXiv:2406.12045.
24. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. 2022. [React: Synergizing reasoning and acting in language models](https://md.research.elamur.ai/free/en/papers/yao-2022-react-synergizing-reasoning-acting-language/). In The eleventh international conference on learning representations.
25. Jeffrey M Zacks and Khena M Swallow. 2007. Event segmentation. Current directions in psychological science, 16(2):80–84. [find](https://scholar.google.com/scholar?q=Jeffrey+M+Zacks+and+Khena+M+Swallow%2E+2007%2E+Event+segmentation%2E+Current+directions+in+psychological+science%2C+16%282%29%3A80%E2%80%9384%2E)
26. Qi Zhao, Haotian Fu, Chen Sun, and George Konidaris. 2024. Epo: Hierarchical llm agents with environment preference optimization. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 6401–6415. [find](https://scholar.google.com/scholar?q=Qi+Zhao%2C+Haotian+Fu%2C+Chen+Sun%2C+and+George+Konidaris%2E+2024%2E+Epo%3A+Hierarchical+llm+agents+with+environment+preference+optimization%2E+In+Proceedings+of+the+2024+Conference+on+Empirical+Methods+in+Natural+)
27. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, and 1 others. 2023. [Judging llm-as-a-judge with mt-bench and chatbot arena](https://md.research.elamur.ai/free/en/papers/zheng-2023-judging-llm-judge-mt-bench/). Advances in neural information processing systems, 36:46595–46623.
28. Shuyan Zhou, Frank F Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, and 1 others. 2023. [Webarena: A realistic web environment for building autonomous agents](https://md.research.elamur.ai/free/en/papers/zhou-2023-webarena-realistic-web-environment-building/). arXiv:2307.13854. arXiv preprint Kunlun Zhu, Zijia Liu, Bingxuan Li, Muxin Tian, Y…
29. Dongqi Zuo, CHEN Zheng, Chuan Zhou, Yandong Guo, Xiao He, and Mingming Gong. 2025. Radi: Llms as world models for robotic action decomposition and imagination. In ICLR 2025 Workshop on World Models: Understanding, Modelling and Scaling. [find](https://scholar.google.com/scholar?q=Dongqi+Zuo%2C+CHEN+Zheng%2C+Chuan+Zhou%2C+Yandong+Guo%2C+Xiao+He%2C+and+Mingming+Gong%2E+2025%2E+Radi%3A+Llms+as+world+models+for+robotic+action+decomposition+and+imagination%2E+In+ICLR+2025+Workshop+on+World+Models%3A+U)

*Updated 1 November 2026*