---
title: "From Individual Excellence to Collective Sustainability: Seeking"
description: "The paper addresses a weakness of proactive LLM agents that coordinate heterogeneous research teams. 25 references."
authors: ["Tong Zhang", "Yang Wu", "Yufei Shi", "Rujing Yao", "Zhuoren Jiang", "Xiaozhong Liu"]
year: 2026
venue: "Findings of the Association for Computational Linguistics: ACL 2026"
source: "https://aclanthology.org/2026.findings-acl.1920/"
doi: "10.18653/v1/2026.findings-acl.1920"
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/zhang-2026-individual-excellence-collective-sustainability-seeking/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/zhang-2026-individual-excellence-collective-sustainability-seeking/)

[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/34/) / From Individual Excellence to Collective Sustainability: Seeking Strategic Equilibrium in Proactive Multi-Agent Teams

# From Individual Excellence to Collective Sustainability: Seeking Strategic Equilibrium in Proactive Multi-Agent Teams

**Authors:** [Tong Zhang](https://md.research.elamur.ai/free/en/authors/tong-zhang/), [Yang Wu](https://md.research.elamur.ai/free/en/authors/yang-wu/), [Yufei Shi](https://md.research.elamur.ai/free/en/authors/yufei-shi/), [Rujing Yao](https://md.research.elamur.ai/free/en/authors/rujing-yao/), [Zhuoren Jiang](https://md.research.elamur.ai/free/en/authors/zhuoren-jiang/) and [Xiaozhong Liu](https://md.research.elamur.ai/free/en/authors/xiaozhong-liu/) · **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.1920/) · [PDF](https://aclanthology.org/2026.findings-acl.1920.pdf) · [DOI 10.18653/v1/2026.findings-acl.1920](https://doi.org/10.18653/v1/2026.findings-acl.1920) · [OpenAlex](https://openalex.org/W7166829474)

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

## Summary
The paper addresses a weakness of proactive LLM agents that coordinate heterogeneous research teams. When they optimize only for immediate task accuracy, they keep sending work to the strongest member, typically the PI, while junior researchers stay idle. The authors call this the Individual-centric Trap. Their framework, GT-PMARL, treats task allocation as a multi-agent game and builds the opportunity cost of each member's time into every decision. It has two layers. The upper layer, trained with GRPO, proposes a varied pool of role–tool–task candidates, filtered by an nnPU (Positive-Unlabeled) scorer that learns from sparse examples of successful coordination. The lower layer uses MARL negotiation under a Nash-Pareto objective to trade off task quality against even workloads, measured with a Gini coefficient. The evaluation uses biomedical, sepsis-related PubMed workflows and a real-world deployment with human teams. GT-PMARL keeps accuracy high while avoiding expert overload. This matters for human-AI collaboration tools that must remain sustainable over time, not just in the short term.

## Key points
- GT-PMARL reaches the highest Social Welfare Value (SWV), 0.2881, with a Gini of 0.3697. Llama-3B has the lowest Gini (0.1455), but its accuracy is poor: it spreads tasks evenly but assigns them wrongly.
- In real-world deployment, GT-PMARL receives a Personalization rating of 4.7778, beating Qwen-plus-base.
- In the ablation study, the 0.3697 score reported for the full model falls to:
  - 0.3172 without GRPO;
  - 0.3399 without the PU-Score;
  - 0.2198 without the multi-agent layer.
  
  The authors conclude that game-theoretic coordination is essential.
- The human evaluation involved three biomedical teams, each with one PI and five PhD students, plus a logistics company team. The data come from public PubMed metadata, and the authors plan to release the curated benchmark.
- Stated limitations:
  - testing is limited to the biomedical domain;
  - proactive planning adds latency;
  - only text input is supported.

*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

- [Zhejiang University](https://md.research.elamur.ai/free/en/universities/zhejiang-university/) — University, Hangzhou, China · 224 works · 13,587 citations
- [Hong Kong Polytechnic University](https://md.research.elamur.ai/free/en/universities/hong-kong-polytechnic-university/) — University, Hong Kong, China · 106 works · 5,190 citations
- [Nankai University](https://md.research.elamur.ai/free/en/universities/nankai-university/) — University, Tianjin, China · 36 works · 762 citations
- [Worcester Polytechnic Institute](https://md.research.elamur.ai/free/en/universities/worcester-polytechnic-institute/) — University, Worcester, United States · 12 works · 380 citations

📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 18,614 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 17,965 works · [💭 Psychology](https://md.research.elamur.ai/free/en/shelves/psychology/) · 1,649 works · [💭 Social Psychology](https://md.research.elamur.ai/free/en/shelves/psychology/social-psychology/) · 330 works

## Cite

Cite as: Tong Zhang, Yang Wu, Yufei Shi, Rujing Yao, Zhuoren Jiang and Xiaozhong Liu. 2026. From Individual Excellence to Collective Sustainability: Seeking Strategic Equilibrium in Proactive Multi-Agent Teams. 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{zhang2026individual,
  title = {{From Individual Excellence to Collective Sustainability: Seeking Strategic Equilibrium in Proactive Multi-Agent Teams}},
  author = {Tong Zhang, Yang Wu, Yufei Shi, Rujing Yao, Zhuoren Jiang and Xiaozhong Liu},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.1920},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.1920}
}
```

- [← previous From Implicit to Explicit: Token-Efficient Logical Supervision for Mathematical Reasoning in LLMs](https://md.research.elamur.ai/free/en/papers/wang-2026-implicit-explicit-token-efficient-logical/)
- [next → From Inheritance to Saturation: Disentangling the Evolution of Visual Redundancy for Architecture-Aware MLLM Inference Acceleration](https://md.research.elamur.ai/free/en/papers/shi-2026-inheritance-saturation-disentangling-evolution-visual/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/34/)

## References

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

1. Sonya Ahamed, Gillian L Galford, Bindu Panikkar, Donna Rizzo, and Jennie C Stephens. 2024. Carbon collusion: Cooperation, competition, and climate obstruction in the global oil and gas extraction network. Energy Policy, 190:114103. [find](https://scholar.google.com/scholar?q=Sonya+Ahamed%2C+Gillian+L+Galford%2C+Bindu+Panikkar%2C+Donna+Rizzo%2C+and+Jennie+C+Stephens%2E+2024%2E+Carbon+collusion%3A+Cooperation%2C+competition%2C+and+climate+obstruction+in+the+global+oil+and+gas+extraction+netw)
2. Kai Barron, Steffen Huck, and Philippe Jehiel. 2024. Everyday econometricians: Selection neglect and overoptimism when learning from others. American Economic Journal: Microeconomics, 16(3):162–198. [find](https://scholar.google.com/scholar?q=Kai+Barron%2C+Steffen+Huck%2C+and+Philippe+Jehiel%2E+2024%2E+Everyday+econometricians%3A+Selection+neglect+and+overoptimism+when+learning+from+others%2E+American+Economic+Journal%3A+Microeconomics%2C+16%283%29%3A162%E2%80%93198%2E)
3. Geng Chen, Xiaoxian Kong, and Qingtian Zeng. 2024. Collaborative localization algorithm for joint node selection and power allocation based on cooperative games. In Proceedings of the 2024 2nd International Conference on Computer, Internet of Things and Smart City, pages 56–61. [find](https://scholar.google.com/scholar?q=Geng+Chen%2C+Xiaoxian+Kong%2C+and+Qingtian+Zeng%2E+2024%2E+Collaborative+localization+algorithm+for+joint+node+selection+and+power+allocation+based+on+cooperative+games%2E+In+Proceedings+of+the+2024+2nd+Interna)
4. Yiqun Chen, Jiaxin Mao, Yi Zhang, Dehong Ma, Long Xia, Jun Fan, Daiting Shi, Zhicong Cheng, Simiu Gu, and Dawei Yin. 2025. Ma4div: Multi-agent reinforcement learning for search result diversification. In Proceedings of the ACM on Web Conference 2025, pages 1703–1715. [find](https://scholar.google.com/scholar?q=Yiqun+Chen%2C+Jiaxin+Mao%2C+Yi+Zhang%2C+Dehong+Ma%2C+Long+Xia%2C+Jun+Fan%2C+Daiting+Shi%2C+Zhicong+Cheng%2C+Simiu+Gu%2C+and+Dawei+Yin%2E+2025%2E+Ma4div%3A+Multi%2Dagent+reinforcement+learning+for+search+result+diversification%2E)
5. Tanner Fiez, Benjamin Chasnov, and Lillian Ratliff. 2020. Implicit learning dynamics in stackelberg games: Equilibria characterization, convergence analysis, and empirical study. In International conference on machine learning, pages 3133–3144. PMLR. [find](https://scholar.google.com/scholar?q=Tanner+Fiez%2C+Benjamin+Chasnov%2C+and+Lillian+Ratliff%2E+2020%2E+Implicit+learning+dynamics+in+stackelberg+games%3A+Equilibria+characterization%2C+convergence+analysis%2C+and+empirical+study%2E+In+International+conf)
6. Rihab Gorsane, Omayma Mahjoub, Ruan John de Kock, Roland Dubb, Siddarth Singh, and Arnu Pretorius. 2022. Towards a standardised performance evaluation protocol for cooperative marl. Advances in Neural Information Processing Systems, 35:5510–5521. [find](https://scholar.google.com/scholar?q=Rihab+Gorsane%2C+Omayma+Mahjoub%2C+Ruan+John+de+Kock%2C+Roland+Dubb%2C+Siddarth+Singh%2C+and+Arnu+Pretorius%2E+2022%2E+Towards+a+standardised+performance+evaluation+protocol+for+cooperative+marl%2E+Advances+in+Neural)
7. Pablo Hernandez-Leal, Michael Kaisers, Tim Baarslag, and Enrique Munoz De Cote. 2017. A survey of learning in multiagent environments: Dealing with non-stationarity. arXiv preprint arXiv:1707.09183. — [arXiv:1707.09183](https://arxiv.org/abs/1707.09183) [find](https://scholar.google.com/scholar?q=Pablo+Hernandez%2DLeal%2C+Michael+Kaisers%2C+Tim+Baarslag%2C+and+Enrique+Munoz+De+Cote%2E+2017%2E+A+survey+of+learning+in+multiagent+environments%3A+Dealing+with+non%2Dstationarity%2E+arXiv+preprint+arXiv%3A1707%2E09183%2E+%E2%80%94)
8. Xinyi Hou, Yanjie Zhao, Shenao Wang, and Haoyu Wang. 2025. Model context protocol (mcp): Landscape, security threats, and future research directions. arXiv preprint arXiv:2503.23278. — [arXiv:2503.23278](https://arxiv.org/abs/2503.23278) [find](https://scholar.google.com/scholar?q=Xinyi+Hou%2C+Yanjie+Zhao%2C+Shenao+Wang%2C+and+Haoyu+Wang%2E+2025%2E+Model+context+protocol+%28mcp%29%3A+Landscape%2C+security+threats%2C+and+future+research+directions%2E+arXiv+preprint+arXiv%3A2503%2E23278%2E+%E2%80%94+%5BarXiv%3A2503%2E232)
9. Kristen Jaskie and Andreas Spanias. 2022. Positive unlabeled learning. Morgan & Claypool Publishers. [find](https://scholar.google.com/scholar?q=Kristen+Jaskie+and+Andreas+Spanias%2E+2022%2E+Positive+unlabeled+learning%2E+Morgan+%26+Claypool+Publishers%2E)
10. Ryuichi Kiryo, Gang Niu, Marthinus C Du Plessis, and Masashi Sugiyama. 2017. Positive-unlabeled learning with non-negative risk estimator. Advances in neural information processing systems, 30. [find](https://scholar.google.com/scholar?q=Ryuichi+Kiryo%2C+Gang+Niu%2C+Marthinus+C+Du+Plessis%2C+and+Masashi+Sugiyama%2E+2017%2E+Positive%2Dunlabeled+learning+with+non%2Dnegative+risk+estimator%2E+Advances+in+neural+information+processing+systems%2C+30%2E)
11. Yanyan Li, Yijun Wang, and Yiwei Zhou. 2024. Multiagent deep reinforcement learning algorithms in starcraft ii: A review. IEEE Access, 12:167452–167470. [find](https://scholar.google.com/scholar?q=Yanyan+Li%2C+Yijun+Wang%2C+and+Yiwei+Zhou%2E+2024%2E+Multiagent+deep+reinforcement+learning+algorithms+in+starcraft+ii%3A+A+review%2E+IEEE+Access%2C+12%3A167452%E2%80%93167470%2E)
12. Bin Lin, Zhenyu Tang, Yang Ye, Jiaxi Cui, Bin Zhu, Peng Jin, Jinfa Huang, Junwu Zhang, Yatian Pang, Munan Ning, et al. 2024. Moe-llava: Mixture of experts for large vision-language models. arXiv preprint arXiv:2401.15947. — [arXiv:2401.15947](https://arxiv.org/abs/2401.15947) [find](https://scholar.google.com/scholar?q=Bin+Lin%2C+Zhenyu+Tang%2C+Yang+Ye%2C+Jiaxi+Cui%2C+Bin+Zhu%2C+Peng+Jin%2C+Jinfa+Huang%2C+Junwu+Zhang%2C+Yatian+Pang%2C+Munan+Ning%2C+et+al%2E+2024%2E+Moe%2Dllava%3A+Mixture+of+experts+for+large+vision%2Dlanguage+models%2E+arXiv+prepr)
13. Zuxin Liu, Thai Hoang, Jianguo Zhang, Ming Zhu, Tian Lan, Juntao Tan, Weiran Yao, Zhiwei Liu, Yihao Feng, Rithesh RN, et al. 2024. Apigen: Automated pipeline for generating verifiable and diverse function-calling datasets. Advances in Neural Information Processing Systems, 37:54463–54482. [find](https://scholar.google.com/scholar?q=Zuxin+Liu%2C+Thai+Hoang%2C+Jianguo+Zhang%2C+Ming+Zhu%2C+Tian+Lan%2C+Juntao+Tan%2C+Weiran+Yao%2C+Zhiwei+Liu%2C+Yihao+Feng%2C+Rithesh+RN%2C+et+al%2E+2024%2E+Apigen%3A+Automated+pipeline+for+generating+verifiable+and+diverse+func)
14. Yaxi Lu, Shenzhi Yang, Cheng Qian, Guirong Chen, Qinyu Luo, Yesai Wu, Huadong Wang, Xin Cong, Zhong Zhang, Yankai Lin, et al. 2024. Proactive agent: Shifting llm agents from reactive responses to active assistance. arXiv preprint arXiv:2410.12361. — [arXiv:2410.12361](https://arxiv.org/abs/2410.12361) [find](https://scholar.google.com/scholar?q=Yaxi+Lu%2C+Shenzhi+Yang%2C+Cheng+Qian%2C+Guirong+Chen%2C+Qinyu+Luo%2C+Yesai+Wu%2C+Huadong+Wang%2C+Xin+Cong%2C+Zhong+Zhang%2C+Yankai+Lin%2C+et+al%2E+2024%2E+Proactive+agent%3A+Shifting+llm+agents+from+reactive+responses+to+acti)
15. Alan O’Sullivan. 2003. Dispersed collaboration in a multi-firm, multi-team product-development project. Journal of Engineering and Technology Management, 20(1-2):93–116. [find](https://scholar.google.com/scholar?q=Alan+O%E2%80%99Sullivan%2E+2003%2E+Dispersed+collaboration+in+a+multi%2Dfirm%2C+multi%2Dteam+product%2Ddevelopment+project%2E+Journal+of+Engineering+and+Technology+Management%2C+20%281%2D2%29%3A93%E2%80%93116%2E)
16. Hugo Roger Paz. 2025. An agent-based simulation of regularity-driven student attrition: How institutional time-to-live constraints create a dropout trap in higher education. arXiv preprint arXiv:2511.16243. — [arXiv:2511.16243](https://arxiv.org/abs/2511.16243) [find](https://scholar.google.com/scholar?q=Hugo+Roger+Paz%2E+2025%2E+An+agent%2Dbased+simulation+of+regularity%2Ddriven+student+attrition%3A+How+institutional+time%2Dto%2Dlive+constraints+create+a+dropout+trap+in+higher+education%2E+arXiv+preprint+arXiv%3A2511%2E)
17. Jiaqi Qin, Yi Zhang, Shixiong Fan, Xiaonan Hu, Yongqiang Huang, Zexin Lu, and Yan Liu. 2022. Multi-task short-term reactive and active load forecasting method based on attention-lstm model. International Journal of Electrical Power & Energy Systems, 135:107517. [find](https://scholar.google.com/scholar?q=Jiaqi+Qin%2C+Yi+Zhang%2C+Shixiong+Fan%2C+Xiaonan+Hu%2C+Yongqiang+Huang%2C+Zexin+Lu%2C+and+Yan+Liu%2E+2022%2E+Multi%2Dtask+short%2Dterm+reactive+and+active+load+forecasting+method+based+on+attention%2Dlstm+model%2E+Internatio)
18. Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, et al. 2023. [Toolllm: Facilitating large language models to master 16000+ real-world apis](https://md.research.elamur.ai/free/en/papers/qin-2023-toolllm-facilitating-large-language-models/). arXiv preprint arXiv:2307.16789.
19. Tianyu Ren, Xuan Yao, Yang Li, and Xiao-Jun Zeng. 2025. Bottom-up reputation promotes cooperation with multi-agent reinforcement learning. arXiv preprint arXiv:2502.01971. — [arXiv:2502.01971](https://arxiv.org/abs/2502.01971) [find](https://scholar.google.com/scholar?q=Tianyu+Ren%2C+Xuan+Yao%2C+Yang+Li%2C+and+Xiao%2DJun+Zeng%2E+2025%2E+Bottom%2Dup+reputation+promotes+cooperation+with+multi%2Dagent+reinforcement+learning%2E+arXiv+preprint+arXiv%3A2502%2E01971%2E+%E2%80%94+%5BarXiv%3A2502%2E01971%5D%28https%3A%2F)
20. 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.
21. Fangzhou Wu, Shutong Wu, Yulong Cao, and Chaowei Xiao. 2024a. Wipi: A new web threat for llm-driven web agents. arXiv preprint arXiv:2402.16965. — [arXiv:2402.16965](https://arxiv.org/abs/2402.16965) [find](https://scholar.google.com/scholar?q=Fangzhou+Wu%2C+Shutong+Wu%2C+Yulong+Cao%2C+and+Chaowei+Xiao%2E+2024a%2E+Wipi%3A+A+new+web+threat+for+llm%2Ddriven+web+agents%2E+arXiv+preprint+arXiv%3A2402%2E16965%2E+%E2%80%94+%5BarXiv%3A2402%2E16965%5D%28https%3A%2F%2Farxiv%2Eorg%2Fabs%2F2402%2E16965%29)
22. Qinzhuo Wu, Wei Liu, Jian Luan, and Bin Wang. 2024b. Toolplanner: A tool augmented llm for multi granularity instructions with path planning and feedback. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 18315–18339. [find](https://scholar.google.com/scholar?q=Qinzhuo+Wu%2C+Wei+Liu%2C+Jian+Luan%2C+and+Bin+Wang%2E+2024b%2E+Toolplanner%3A+A+tool+augmented+llm+for+multi+granularity+instructions+with+path+planning+and+feedback%2E+In+Proceedings+of+the+2024+Conference+on+Empi)
23. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. [React: Synergizing reasoning and acting in language models](https://md.research.elamur.ai/free/en/papers/yao-2022-react-synergizing-reasoning-acting-language/). In International Conference on Learning Representations (ICLR).
24. Hongyi Yuan, Zheng Yuan, Chuanqi Tan, Wei Wang, Songfang Huang, and Fei Huang. 2023. Rrhf: Rank responses to align language models with human feedback. Advances in Neural Information Processing Systems, 36:10935–10950. [find](https://scholar.google.com/scholar?q=Hongyi+Yuan%2C+Zheng+Yuan%2C+Chuanqi+Tan%2C+Wei+Wang%2C+Songfang+Huang%2C+and+Fei+Huang%2E+2023%2E+Rrhf%3A+Rank+responses+to+align+language+models+with+human+feedback%2E+Advances+in+Neural+Information+Processing+System)
25. Ceyao Zhang, Kaijie Yang, Siyi Hu, Zihao Wang, Guanghe Li, Yihang Sun, Cheng Zhang, Zhaowei Zhang, Anji Liu, Song-Chun Zhu, et al. 2024. Proagent: building proactive cooperative agents with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 1759… [find](https://scholar.google.com/scholar?q=Ceyao+Zhang%2C+Kaijie+Yang%2C+Siyi+Hu%2C+Zihao+Wang%2C+Guanghe+Li%2C+Yihang+Sun%2C+Cheng+Zhang%2C+Zhaowei+Zhang%2C+Anji+Liu%2C+Song%2DChun+Zhu%2C+et+al%2E+2024%2E+Proagent%3A+building+proactive+cooperative+agents+with+large+lang)

*Updated 2 November 2026*