---
title: "Do We Always Need Query-Level Workflows? Rethinking Agentic"
description: "Their framework, SCALE, replaces most of that evaluation: the LLM optimizer predicts each candidate's… 0.61% accuracy. 21 references."
authors: ["Zixu Wang", "Bingbing Xu", "Yige Yuan", "Huawei Shen", "Xueqi Cheng"]
year: 2026
venue: "Findings of the Association for Computational Linguistics: ACL 2026"
source: "https://aclanthology.org/2026.findings-acl.254/"
doi: "10.18653/v1/2026.findings-acl.254"
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-do-we-always-need-query/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/wang-2026-do-we-always-need-query/)

[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/27/) / Do We Always Need Query-Level Workflows? Rethinking Agentic Workflow Generation for Multi-Agent Systems

# Do We Always Need Query-Level Workflows? Rethinking Agentic Workflow Generation for Multi-Agent Systems

**Authors:** Zixu Wang, [Bingbing Xu](https://md.research.elamur.ai/free/en/authors/bingbing-xu/), [Yige Yuan](https://md.research.elamur.ai/free/en/authors/yige-yuan/), [Huawei Shen](https://md.research.elamur.ai/free/en/authors/huawei-shen/) and [Xueqi Cheng](https://md.research.elamur.ai/free/en/authors/xueqi-cheng/) · **Year:** 2026 · **Venue:** [Findings of the Association for Computational Linguistics: ACL 2026](https://md.research.elamur.ai/free/en/venues/findings/) ([Findings of the Association for Computational Linguistics 2026](https://md.research.elamur.ai/free/en/venues/findings/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.findings-acl.254/) · [PDF](https://aclanthology.org/2026.findings-acl.254.pdf) · [DOI 10.18653/v1/2026.findings-acl.254](https://doi.org/10.18653/v1/2026.findings-acl.254) · [OpenAlex](https://openalex.org/W7166866414)

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

## Summary
LLM-based multi-agent systems coordinate their agents through workflows. These workflows are built either once per task or separately for each query. The authors ask whether per-query generation is really needed. They also ask whether task-level search must pay for a full validation run on every candidate. On DROP, HumanEval, GSM8K and MATH, they compare Aflow's task-level workflows (top-1, top-5, and five repeated runs of the top-1) with the query-level method ScoreFlow. A few reusable task-level workflows, or even repeated runs of a single one, solve about as many test queries as per-query generation. The authors also find that full-execution evaluation is very expensive and often an unreliable guide for the search. Their framework, SCALE, replaces most of that evaluation: the LLM optimizer predicts each candidate's score itself, and a few real executions calibrate the prediction. Across six benchmarks, accuracy drops only slightly compared with Aflow, while token consumption falls sharply. This makes workflow optimization far cheaper to run.

## Key points
- On four benchmarks, Aflow's full generation process uses roughly 10^6–10^8 LLM tokens, and candidate evaluation dominates that cost.
- The union of Aflow's top-5 workflows covers more test queries than ScoreFlow's per-query workflows reach in accuracy. Five repeated runs of the top-1 workflow give comparable coverage, so much of the query-level gain can be explained by execution randomness.
- Compared with Aflow, SCALE loses 0.61% accuracy on average and cuts total tokens by 54–83%. It also uses fewer tokens than AgentPrune at similar accuracy.
- The calibrated surrogate score S_b tracks real execution scores with MAE 0.16 and Pearson 0.52. Self-prediction alone has low error but near-zero correlation. Few-shot execution alone correlates better but has a much larger error. Confidence-based scoring fails.
- Code: github.com/Camel-Prince/SCALE. Cross-domain generalization was not tested.

*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

- [Chinese Academy of Sciences](https://md.research.elamur.ai/free/en/government/chinese-academy-of-sciences/) — Government body, Beijing, China · 194 works · 23,906 citations
- [Institute of Computing Technology](https://md.research.elamur.ai/free/en/labs/institute-of-computing-technology/) — Research institute, Beijing, China · 51 works · 3,411 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: Zixu Wang, Bingbing Xu, Yige Yuan, Huawei Shen and Xueqi Cheng. 2026. Do We Always Need Query-Level Workflows? Rethinking Agentic Workflow Generation for Multi-Agent Systems. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@inproceedings{wang2026do,
  title = {{Do We Always Need Query-Level Workflows? Rethinking Agentic Workflow Generation for Multi-Agent Systems}},
  author = {Zixu Wang, Bingbing Xu, Yige Yuan, Huawei Shen and Xueqi Cheng},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.254},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.254}
}
```

- [← previous Do Syntactic Features Help Biomedical Relation Extraction? An Empirical Study of Verb Token and Dependency Graph Augmentation](https://md.research.elamur.ai/free/en/papers/sikder-2026-do-syntactic-features-help-biomedical/)
- [next → Doaa Sulaiman at AR-MS NakbaNLP 2026: Faithful Diplomatic Transcription of Arabic Manuscripts Using a Human-Centred Annotation Framework](https://md.research.elamur.ai/free/en/papers/sulaiman-2026-doaa-sulaiman-ar-ms-nakbanlp/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/27/)

## References

21 references, 13 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie Cai, Michael Terry, Quoc Le, and 1 others. 2021. [Program synthesis with large language models](https://md.research.elamur.ai/free/en/papers/austin-2021-program-synthesis-large-language-models/). arXiv preprint arXiv:2108.07732.
2. Mark Chen. 2021. [Evaluating large language models trained on code](https://md.research.elamur.ai/free/en/papers/chen-2021-evaluating-large-language-models-trained/). arXiv preprint arXiv:2107.03374.
3. Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, and 1 others. 2021. [Training verifiers to solve math word problems](https://md.research.elamur.ai/free/en/papers/cobbe-2021-training-verifiers-solve-math-word/). arXiv preprint arXiv:2110.14168.
4. Dheeru Dua, Yizhong Wang, Pradeep Dasigi, Gabriel Stanovsky, Sameer Singh, and Matt Gardner. 2019. [Drop: A reading comprehension benchmark requiring discrete reasoning over paragraphs](https://md.research.elamur.ai/free/en/papers/dua-2019-drop-reading-comprehension-benchmark-requiring/). arXiv preprint arXiv:1903.00161.
5. Hongcheng Gao, Yue Liu, Yufei He, Longxu Dou, Chao Du, Zhijie Deng, Bryan Hooi, Min Lin, and Tianyu Pang. 2025. Flowreasoner: Reinforcing query-level meta-agents. arXiv preprint arXiv:2504.15257. — [arXiv:2504.15257](https://arxiv.org/abs/2504.15257) [find](https://scholar.google.com/scholar?q=Hongcheng+Gao%2C+Yue+Liu%2C+Yufei+He%2C+Longxu+Dou%2C+Chao+Du%2C+Zhijie+Deng%2C+Bryan+Hooi%2C+Min+Lin%2C+and+Tianyu+Pang%2E+2025%2E+Flowreasoner%3A+Reinforcing+query%2Dlevel+meta%2Dagents%2E+arXiv+preprint+arXiv%3A2504%2E15257%2E+%E2%80%94+%5Ba)
6. Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. 2021. [Measuring mathematical problem solving with the math dataset](https://md.research.elamur.ai/free/en/papers/hendrycks-2021-measuring-mathematical-problem-solving-math/). arXiv preprint arXiv:2103.03874.
7. Shengran Hu, Cong Lu, and Jeff Clune. 2024. Automated design of agentic systems. arXiv preprint arXiv:2408.08435. [find](https://scholar.google.com/scholar?q=Shengran+Hu%2C+Cong+Lu%2C+and+Jeff+Clune%2E+2024%2E+Automated+design+of+agentic+systems%2E+arXiv+preprint+arXiv%3A2408%2E08435%2E)
8. Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jiajun Zhang, Bowen Yu, Keming Lu, and 1 others. 2024. Qwen2. 5-coder technical report. arXiv preprint arXiv:2409.12186. — [Qwen2.5-Coder Technical Report](https://md.research.elamur.ai/free/en/papers/hui-2024-qwen2-5-coder-technical-report/)
9. Tian Liang, Zhiwei He, Wenxiang Jiao, Xing Wang, Yan Wang, Rui Wang, Yujiu Yang, Shuming Shi, and Zhaopeng Tu. 2024. [Encouraging divergent thinking in large language models through multi-agent debate](https://md.research.elamur.ai/free/en/papers/liang-2023-encouraging-divergent-thinking-large-language/). In Proceedings of the 2024 conference on empirical methods in natural language processing, pages 17…
10. Boye Niu, Yiliao Song, Kai Lian, Yifan Shen, Yu Yao, Kun Zhang, and Tongliang Liu. 2025. Flow: Modularized agentic workflow automation. arXiv preprint arXiv:2501.07834. — [arXiv:2501.07834](https://arxiv.org/abs/2501.07834) [find](https://scholar.google.com/scholar?q=Boye+Niu%2C+Yiliao+Song%2C+Kai+Lian%2C+Yifan+Shen%2C+Yu+Yao%2C+Kun+Zhang%2C+and+Tongliang+Liu%2E+2025%2E+Flow%3A+Modularized+agentic+workflow+automation%2E+arXiv+preprint+arXiv%3A2501%2E07834%2E+%E2%80%94+%5BarXiv%3A2501%2E07834%5D%28https%3A%2F%2Far)
11. Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. [Direct preference optimization: Your language model is secretly a reward model](https://md.research.elamur.ai/free/en/papers/rafailov-2023-direct-preference-optimization-language-model/). Advances in neural information processing systems, 36:53728–53741.
12. 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.
13. Yinjie Wang, Ling Yang, Guohao Li, Mengdi Wang, and Bryon Aragam. 2025a. Scoreflow: Mastering llm agent workflows via score-based preference optimization. arXiv preprint arXiv:2502.04306. — [arXiv:2502.04306](https://arxiv.org/abs/2502.04306) [find](https://scholar.google.com/scholar?q=Yinjie+Wang%2C+Ling+Yang%2C+Guohao+Li%2C+Mengdi+Wang%2C+and+Bryon+Aragam%2E+2025a%2E+Scoreflow%3A+Mastering+llm+agent+workflows+via+score%2Dbased+preference+optimization%2E+arXiv+preprint+arXiv%3A2502%2E04306%2E+%E2%80%94+%5BarXiv%3A250)
14. Zhexuan Wang, Yutong Wang, Xuebo Liu, Liang Ding, Miao Zhang, Jie Liu, and Min Zhang. 2025b. [Agentdropout: Dynamic agent elimination for token-efficient and high-performance llm-based multi-agent collaboration](https://md.research.elamur.ai/free/en/papers/wang-2025-agentdropout-dynamic-agent-elimination-token/). arXiv preprint arXiv:2503.18891.
15. Shengxiang Xu, Jiayi Zhang, Shimin Di, Yuyu Luo, Liang Yao, Hanmo Liu, Jia Zhu, Fan Liu, and Min-Ling Zhang. 2025. Robustflow: Towards robust agentic workflow generation. arXiv preprint arXiv:2509.21834. — [arXiv:2509.21834](https://arxiv.org/abs/2509.21834) [find](https://scholar.google.com/scholar?q=Shengxiang+Xu%2C+Jiayi+Zhang%2C+Shimin+Di%2C+Yuyu+Luo%2C+Liang+Yao%2C+Hanmo+Liu%2C+Jia+Zhu%2C+Fan+Liu%2C+and+Min%2DLing+Zhang%2E+2025%2E+Robustflow%3A+Towards+robust+agentic+workflow+generation%2E+arXiv+preprint+arXiv%3A2509%2E218)
16. 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](https://md.research.elamur.ai/free/en/papers/yang-2025-qwen3-technical-report/). arXiv preprint arXiv:2505.09388.
17. Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D Manning. 2018. [Hotpotqa: A dataset for diverse, explainable multi-hop question answering](https://md.research.elamur.ai/free/en/papers/yang-2018-hotpotqa-dataset-diverse-explainable-multi/). In Proceedings of the 2018 conference on empirical methods in natural language processing, pages 2369–…
18. 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. 5157 9
19. Rui Ye, Shuo Tang, Rui Ge, Yaxin Du, Zhenfei Yin, Siheng Chen, and Jing Shao. 2025. Mas-gpt: Training llms to build llm-based multi-agent systems. arXiv preprint arXiv:2503.03686. which represents your estimation of its overall accuracy . 6 Respond with a brief explanation first , followed by a sing… — [arXiv:2503.03686](https://arxiv.org/abs/2503.03686) [find](https://scholar.google.com/scholar?q=Rui+Ye%2C+Shuo+Tang%2C+Rui+Ge%2C+Yaxin+Du%2C+Zhenfei+Yin%2C+Siheng+Chen%2C+and+Jing+Shao%2E+2025%2E+Mas%2Dgpt%3A+Training+llms+to+build+llm%2Dbased+multi%2Dagent+systems%2E+arXiv+preprint+arXiv%3A2503%2E03686%2E+which+represents+you)
20. Guibin Zhang, Luyang Niu, Junfeng Fang, Kun Wang, 7 Lei Bai, and Xiang Wang. 2025. Multi-agent archi- 8 Dataset Description : tecture search via agentic supernet. arXiv preprint 9 \\< dataset \> 10 { dataset_description } arXiv:2502.04180. 11 \\</ dataset \> Guibin Zhang, Yanwei Yue, Zhixun Li, Sukwon Yun… [find](https://scholar.google.com/scholar?q=Guibin+Zhang%2C+Luyang+Niu%2C+Junfeng+Fang%2C+Kun+Wang%2C+7+Lei+Bai%2C+and+Xiang+Wang%2E+2025%2E+Multi%2Dagent+archi%2D+8+Dataset+Description+%3A+tecture+search+via+agentic+supernet%2E+arXiv+preprint+9+%5C%3C+dataset+%5C%3E+10+%7B+d)
21. Mingchen Zhuge, Wenyi Wang, Louis Kirsch, 24 \\</ prompt \> Francesco Faccio, Dmitrii Khizbullin, and Jürgen 25 Schmidhuber. 2024. Gptswarm: Language agents 26 The workflow is Python code ; the key function is __call__ ( question ) , as optimizable graphs. In Forty-first International which produces th… [find](https://scholar.google.com/scholar?q=Mingchen+Zhuge%2C+Wenyi+Wang%2C+Louis+Kirsch%2C+24+%5C%3C%2F+prompt+%5C%3E+Francesco+Faccio%2C+Dmitrii+Khizbullin%2C+and+J%C3%BCrgen+25+Schmidhuber%2E+2024%2E+Gptswarm%3A+Language+agents+26+The+workflow+is+Python+code+%3B+the+key+fun)

*Updated 1 November 2026*