---
title: "Agent-Dice — paper by Wu et al. (2026) — summary and ideas"
description: "The authors support the design with theory: a first-order linearization argument, and a sign-noise model in which excluding… 60 references."
authors: ["Zheng Wu", "Xingyu Lou", "Xinbei Ma", "Yansi Li", "Weiwen Liu", "Weinan Zhang", "Jun Wang", "Zhuosheng Zhang"]
year: 2026
venue: "Findings of the Association for Computational Linguistics: ACL 2026"
source: "https://aclanthology.org/2026.findings-acl.908/"
doi: "10.18653/v1/2026.findings-acl.908"
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/wu-2026-agent-dice-disentangling-knowledge-updates/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/wu-2026-agent-dice-disentangling-knowledge-updates/)

[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/) / Agent-Dice: Disentangling Knowledge Updates via Geometric Consensus for Agent Continual Learning

# Agent-Dice: Disentangling Knowledge Updates via Geometric Consensus for Agent Continual Learning

**Authors:** Zheng Wu, [Xingyu Lou](https://md.research.elamur.ai/free/en/authors/xingyu-lou/), [Xinbei Ma](https://md.research.elamur.ai/free/en/authors/xinbei-ma/), Yansi Li, [Weiwen Liu](https://md.research.elamur.ai/free/en/authors/weiwen-liu/), [Weinan Zhang](https://md.research.elamur.ai/free/en/authors/weinan-zhang/), [Jun Wang](https://md.research.elamur.ai/free/en/authors/jun-wang/) and [Zhuosheng Zhang](https://md.research.elamur.ai/free/en/authors/zhuosheng-zhang/) · **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.908/) · [PDF](https://aclanthology.org/2026.findings-acl.908.pdf) · [DOI 10.18653/v1/2026.findings-acl.908](https://doi.org/10.18653/v1/2026.findings-acl.908) · [OpenAlex](https://openalex.org/W7166850039)

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

## Summary

The paper addresses continual learning for LLM-based agents, where acquiring new tasks tends to erase earlier skills (the stability–plasticity dilemma). The authors attribute this to methods treating all knowledge updates alike instead of separating knowledge shared across tasks from task-specific updates that conflict. Agent-Dice is a parameter-fusion framework. It takes task vectors (each fine-tuned model minus the pre-trained weights) and merges them back onto the base agent in two steps. First, geometric consensus filtering discards parameter updates whose sign disagrees with the majority direction. Second, curvature-based importance weighting boosts directions that tasks agree on. The authors support the design with theory: a first-order linearization argument, and a sign-noise model in which excluding outlier tasks makes error probability shrink exponentially with consensus size. Experiments cover GUI agents and tool-use agents. Agent-Dice improves continual-learning performance over sequential training and common merging baselines, needs only minutes of extra compute, and changes parameters only modestly. This makes it a cheap, training-free option for keeping agents up to date.

## Key points

- **Data and models.** GUI-agent experiments use AITZ, AndroidControl and GUI-Odyssey with OS-Atlas-Pro-7B. Tool-use experiments learn from four data subsets (Subset 0–3) with Qwen3-8B.
- **Comparison with merging methods (GUI domain, Table 6).** Agent-Dice reaches 51.42 on AndroidControl and 72.28 on GUI-Odyssey, the best of the four methods on both. On AITZ it scores 57.10, slightly below Adapter-soups (57.67). The other baselines are Model soups and AdapterFusion.
- **Ablations.** Removing either Stage 1 (consensus filtering) or Stage 2 (importance weighting) lowers performance in both domains.
- **Model similarity (KL / L2 heatmaps).**
  - In the tool-use domain, the merged model stays closest to base Qwen3-8B (L2 = 10.80).
  - In the GUI domain, its distance from OS-Atlas-Pro-7B is only marginally above that of single-dataset models.
- **Overhead.** Fusion takes about 62–89 seconds on GPU and roughly ten minutes on CPU. This is negligible compared with hours of training.

*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

- [Shanghai Jiao Tong University](https://md.research.elamur.ai/free/en/universities/shanghai-jiao-tong-university/) — University, Shanghai, China · 212 works · 19,544 citations

📚 Shelves (2): [🖥️ 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

## Cite

Cite as: Zheng Wu, Xingyu Lou, Xinbei Ma, Yansi Li, Weiwen Liu, Weinan Zhang, Jun Wang and Zhuosheng Zhang. 2026. Agent-Dice: Disentangling Knowledge Updates via Geometric Consensus for Agent Continual Learning. 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{wu2026agent,
  title = {{Agent-Dice: Disentangling Knowledge Updates via Geometric Consensus for Agent Continual Learning}},
  author = {Zheng Wu, Xingyu Lou, Xinbei Ma, Yansi Li, Weiwen Liu, Weinan Zhang, Jun Wang and Zhuosheng Zhang},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.908},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.908}
}
```

- [← previous Agent Newsroom: Efficient Chronological Report Generation via Dynamic Multi-Agent Collaboration](https://md.research.elamur.ai/free/en/papers/wang-2026-agent-newsroom-efficient-chronological-report/)
- [next → Agent-GWO: Collaborative Agents for Dynamic Prompt Optimization in Large Language Models](https://md.research.elamur.ai/free/en/papers/wang-2026-agent-gwo-collaborative-agents-dynamic/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/21/)

## References

60 references, 17 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Hao Bai, Yifei Zhou, Jiayi Pan, Mert Cemri, Alane Suhr, Sergey Levine, and Aviral Kumar. 2024. Digirl: Training in-the-wild device-control agents with autonomous reinforcement learning. Advances in Neural Information Processing Systems, 37:12461– 12495. [find](https://scholar.google.com/scholar?q=Hao+Bai%2C+Yifei+Zhou%2C+Jiayi+Pan%2C+Mert+Cemri%2C+Alane+Suhr%2C+Sergey+Levine%2C+and+Aviral+Kumar%2E+2024%2E+Digirl%3A+Training+in%2Dthe%2Dwild+device%2Dcontrol+agents+with+autonomous+reinforcement+learning%2E+Advances+in+Ne)
2. Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, Wenbin Ge, Zhifang Guo, Qidong Huang, Jie Huang, Fei Huang, Binyuan Hui, Shutong Jiang, Zhaohai Li, Mingsheng Li, and 45 others. 2025. [Qwen3-vl technical report](https://md.research.elamur.ai/free/en/papers/bai-2025-qwen3-vl-technical-report/). arXiv prepri…
3. Victor Barres, Honghua Dong, Soham Ray, Xujie Si, and Karthik Narasimhan. 2025. [τ 2 -bench: Evaluating conversational agents in a dual-control environment](https://md.research.elamur.ai/free/en/papers/barres-2025-tau-2-bench-evaluating-conversational/). Preprint, arXiv:2506.07982.
4. Chen Chen, Xinlong Hao, Weiwen Liu, Xu Huang, Xingshan Zeng, Shuai Yu, Dexun Li, Shuai Wang, Weinan Gan, Yuefeng Huang, Wulong Liu, Xinzhi Wang, Defu Lian, Baoqun Yin, Yasheng Wang, and Wu Liu. 2025. Acebench: Who wins the match point in tool usage? Preprint, arXiv:2501.12851. — [arXiv:2501.12851](https://arxiv.org/abs/2501.12851) [find](https://scholar.google.com/scholar?q=Chen+Chen%2C+Xinlong+Hao%2C+Weiwen+Liu%2C+Xu+Huang%2C+Xingshan+Zeng%2C+Shuai+Yu%2C+Dexun+Li%2C+Shuai+Wang%2C+Weinan+Gan%2C+Yuefeng+Huang%2C+Wulong+Liu%2C+Xinzhi+Wang%2C+Defu+Lian%2C+Baoqun+Yin%2C+Yasheng+Wang%2C+and+Wu+Liu%2E+2025%2E+)
5. Yanyu Chen, Jiyue Jiang, Jiahong Liu, Yifei Zhang, Xiao Guo, and Irwin King. 2026. Trace: Trajectory-aware comprehensive evaluation for deep research agents. arXiv preprint arXiv:2602.21230. — [arXiv:2602.21230](https://arxiv.org/abs/2602.21230) [find](https://scholar.google.com/scholar?q=Yanyu+Chen%2C+Jiyue+Jiang%2C+Jiahong+Liu%2C+Yifei+Zhang%2C+Xiao+Guo%2C+and+Irwin+King%2E+2026%2E+Trace%3A+Trajectory%2Daware+comprehensive+evaluation+for+deep+research+agents%2E+arXiv+preprint+arXiv%3A2602%2E21230%2E+%E2%80%94+%5BarXiv%3A)
6. Alexandra Chronopoulou, Matthew E Peters, Alexander Fraser, and Jesse Dodge. 2023. Adaptersoup: Weight averaging to improve generalization of pretrained language models. In Findings of the Association for Computational Linguistics: EACL 2023, pages 2054–2063. [find](https://scholar.google.com/scholar?q=Alexandra+Chronopoulou%2C+Matthew+E+Peters%2C+Alexander+Fraser%2C+and+Jesse+Dodge%2E+2023%2E+Adaptersoup%3A+Weight+averaging+to+improve+generalization+of+pretrained+language+models%2E+In+Findings+of+the+Association)
7. Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, and 1 others. 2024. [The llama 3 herd of models](https://md.research.elamur.ai/free/en/papers/grattafiori-2024-llama-3-herd-models/). arXiv e-prints, pages arXiv–2407.
8. Jinyuan Fang, Yanwen Peng, Xi Zhang, Yingxu Wang, Xinhao Yi, Guibin Zhang, Yi Xu, Bin Wu, Siwei Liu, Zihao Li, and 1 others. 2025. A comprehensive survey of self-evolving ai agents: A new paradigm bridging foundation models and lifelong agentic systems. arXiv preprint arXiv:2508.07407. — [arXiv:2508.07407](https://arxiv.org/abs/2508.07407) [find](https://scholar.google.com/scholar?q=Jinyuan+Fang%2C+Yanwen+Peng%2C+Xi+Zhang%2C+Yingxu+Wang%2C+Xinhao+Yi%2C+Guibin+Zhang%2C+Yi+Xu%2C+Bin+Wu%2C+Siwei+Liu%2C+Zihao+Li%2C+and+1+others%2E+2025%2E+A+comprehensive+survey+of+self%2Devolving+ai+agents%3A+A+new+paradigm+bri)
9. Huan-ang Gao, Jiayi Geng, Wenyue Hua, Mengkang Hu, Xinzhe Juan, Hongzhang Liu, Shilong Liu, Jiahao Qiu, Xuan Qi, Yiran Wu, and 1 others. 2025. A survey of self-evolving agents: On path to artificial super intelligence. arXiv preprint arXiv:2507.21046. [find](https://scholar.google.com/scholar?q=Huan%2Dang+Gao%2C+Jiayi+Geng%2C+Wenyue+Hua%2C+Mengkang+Hu%2C+Xinzhe+Juan%2C+Hongzhang+Liu%2C+Shilong+Liu%2C+Jiahao+Qiu%2C+Xuan+Qi%2C+Yiran+Wu%2C+and+1+others%2E+2025%2E+A+survey+of+self%2Devolving+agents%3A+On+path+to+artificial+s)
10. Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. 2019. Learning a unified classifier incrementally via rebalancing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 831–839. [find](https://scholar.google.com/scholar?q=Saihui+Hou%2C+Xinyu+Pan%2C+Chen+Change+Loy%2C+Zilei+Wang%2C+and+Dahua+Lin%2E+2019%2E+Learning+a+unified+classifier+incrementally+via+rebalancing%2E+In+Proceedings+of+the+IEEE%2FCVF+conference+on+computer+vision+and+p)
11. Jincai Huang, Yongjun Xu, Qi Wang, Qi Cheems Wang, Xingxing Liang, Fei Wang, Zhao Zhang, Wei Wei, Boxuan Zhang, Libo Huang, and 1 others. 2025. Foundation models and intelligent decision-making: Progress, challenges, and perspectives. The Innovation. [find](https://scholar.google.com/scholar?q=Jincai+Huang%2C+Yongjun+Xu%2C+Qi+Wang%2C+Qi+Cheems+Wang%2C+Xingxing+Liang%2C+Fei+Wang%2C+Zhao+Zhang%2C+Wei+Wei%2C+Boxuan+Zhang%2C+Libo+Huang%2C+and+1+others%2E+2025%2E+Foundation+models+and+intelligent+decision%2Dmaking%3A+Progr)
12. Xu Huang, Weiwen Liu, Xiaolong Chen, Xingmei Wang, Hao Wang, Defu Lian, Yasheng Wang, Ruiming Tang, and Enhong Chen. 2024. [Understanding the planning of llm agents: A survey](https://md.research.elamur.ai/free/en/papers/huang-2024-understanding-planning-llm-agents-survey/). arXiv preprint arXiv:2402.02716.
13. Jiyue Jiang, Yanyu Chen, Pengan Chen, Kai Liu, Jingqi Zhou, Zheyong Zhu, He Hu, Fei Ma, Qi Tian, and Chuan Wu. 2026a. A principle-driven adaptive policy for group cognitive stimulation dialogue for elderly with cognitive impairment. [find](https://scholar.google.com/scholar?q=Jiyue+Jiang%2C+Yanyu+Chen%2C+Pengan+Chen%2C+Kai+Liu%2C+Jingqi+Zhou%2C+Zheyong+Zhu%2C+He+Hu%2C+Fei+Ma%2C+Qi+Tian%2C+and+Chuan+Wu%2E+2026a%2E+A+principle%2Ddriven+adaptive+policy+for+group+cognitive+stimulation+dialogue+for+el)
14. Jiyue Jiang, Yanyu Chen, Qingchuan Zhang, Jiayi Li, Xiangyu Shi, Chang Zhou, Ziqian Lin, Jiuming Wang, Dongchen He, Liang Hong, and 1 others. 2026b. Rmsagen: Integrating multiple sequence alignment for function rna design. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40,… [find](https://scholar.google.com/scholar?q=Jiyue+Jiang%2C+Yanyu+Chen%2C+Qingchuan+Zhang%2C+Jiayi+Li%2C+Xiangyu+Shi%2C+Chang+Zhou%2C+Ziqian+Lin%2C+Jiuming+Wang%2C+Dongchen+He%2C+Liang+Hong%2C+and+1+others%2E+2026b%2E+Rmsagen%3A+Integrating+multiple+sequence+alignment+fo)
15. Jiyue Jiang, Alfred Kar Yin Truong, Yanyu Chen, Qinghang Bao, Sheng Wang, Pengan Chen, Jiuming Wang, Lingpeng Kong, Yu Li, and Chuan Wu. 2025. Developing and utilizing a large-scale cantonese dataset for multi-tasking in large language models. In Findings of the Association for Computational Linguis… [find](https://scholar.google.com/scholar?q=Jiyue+Jiang%2C+Alfred+Kar+Yin+Truong%2C+Yanyu+Chen%2C+Qinghang+Bao%2C+Sheng+Wang%2C+Pengan+Chen%2C+Jiuming+Wang%2C+Lingpeng+Kong%2C+Yu+Li%2C+and+Chuan+Wu%2E+2025%2E+Developing+and+utilizing+a+large%2Dscale+cantonese+dataset+)
16. James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, and 1 others. 2017. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(1… [find](https://scholar.google.com/scholar?q=James+Kirkpatrick%2C+Razvan+Pascanu%2C+Neil+Rabinowitz%2C+Joel+Veness%2C+Guillaume+Desjardins%2C+Andrei+A+Rusu%2C+Kieran+Milan%2C+John+Quan%2C+Tiago+Ramalho%2C+Agnieszka+Grabska%2DBarwinska%2C+and+1+others%2E+2017%2E+Overcomin)
17. Songze Li, Xiaoke Guo, Tianqi Liu, Biao Yi, Zhaoyan Gong, Zhiqiang Liu, Huajun Chen, and Wen Zhang. 2026. What’s missing in screen-to-action? towards a ui-in-the-loop paradigm for multimodal gui reasoning. arXiv preprint arXiv:2604.06995. 2021. Linear mode connectivity in multitask and continual lea… [find](https://scholar.google.com/scholar?q=Songze+Li%2C+Xiaoke+Guo%2C+Tianqi+Liu%2C+Biao+Yi%2C+Zhaoyan+Gong%2C+Zhiqiang+Liu%2C+Huajun+Chen%2C+and+Wen+Zhang%2E+2026%2E+What%E2%80%99s+missing+in+screen%2Dto%2Daction%3F+towards+a+ui%2Din%2Dthe%2Dloop+paradigm+for+multimodal+gui+reaso)
18. Wei Li, William E Bishop, Alice Li, Christopher Rawles, Folawiyo Campbell-Ajala, Divya Tyamagundlu, and Oriana Riva. 2024. [On the effects of data scale on ui control agents](https://md.research.elamur.ai/free/en/papers/li-2024-effects-data-scale-ui-control/). Advances in Neural Information Processing Systems, 37:92130–92154.
19. Aneesh Muppidi, Zhiyu Zhang, and Heng Yang. 2024. Fast trac: A parameter-free optimizer for lifelong reinforcement learning. Advances in Neural Information Processing Systems, 37:51169–51195. [find](https://scholar.google.com/scholar?q=Aneesh+Muppidi%2C+Zhiyu+Zhang%2C+and+Heng+Yang%2E+2024%2E+Fast+trac%3A+A+parameter%2Dfree+optimizer+for+lifelong+reinforcement+learning%2E+Advances+in+Neural+Information+Processing+Systems%2C+37%3A51169%E2%80%9351195%2E)
20. Xinzhe Li. 2025. A review of prominent paradigms for llm-based agents: Tool use, planning (including rag), and feedback learning. In Proceedings of the 31st International Conference on Computational Linguistics, pages 9760–9779. [find](https://scholar.google.com/scholar?q=Xinzhe+Li%2E+2025%2E+A+review+of+prominent+paradigms+for+llm%2Dbased+agents%3A+Tool+use%2C+planning+%28including+rag%29%2C+and+feedback+learning%2E+In+Proceedings+of+the+31st+International+Conference+on+Computational+L)
21. Siru Ouyang, Jun Yan, I Hsu, Yanfei Chen, Ke Jiang, Zifeng Wang, Rujun Han, Long T Le, Samira Daruki, Xiangru Tang, and 1 others. 2025. Reasoningbank: Scaling agent self-evolving with reasoning memory. arXiv preprint arXiv:2509.25140. — [arXiv:2509.25140](https://arxiv.org/abs/2509.25140) [find](https://scholar.google.com/scholar?q=Siru+Ouyang%2C+Jun+Yan%2C+I+Hsu%2C+Yanfei+Chen%2C+Ke+Jiang%2C+Zifeng+Wang%2C+Rujun+Han%2C+Long+T+Le%2C+Samira+Daruki%2C+Xiangru+Tang%2C+and+1+others%2E+2025%2E+Reasoningbank%3A+Scaling+agent+self%2Devolving+with+reasoning+memory)
22. Weiwen Liu, Xu Huang, Xingshan Zeng, Shuai Yu, Dexun Li, Shuai Wang, Weinan Gan, Zhengying Liu, Yuanqing Yu, Zezhong WANG, and 1 others. 2025a. [Toolace: Winning the points of llm function calling](https://md.research.elamur.ai/free/en/papers/liu-2024-toolace-winning-points-llm-function/). In The Thirteenth International Conference on Learning Representations.
23. Shishir G Patil, Huanzhi Mao, Fanjia Yan, Charlie Cheng-Jie Ji, Vishnu Suresh, Ion Stoica, and Joseph E. Gonzalez. 2025. The berkeley function calling leaderboard (BFCL): From tool use to agentic evaluation of large language models. In Forty-second International Conference on Machine Learning. [find](https://scholar.google.com/scholar?q=Shishir+G+Patil%2C+Huanzhi+Mao%2C+Fanjia+Yan%2C+Charlie+Cheng%2DJie+Ji%2C+Vishnu+Suresh%2C+Ion+Stoica%2C+and+Joseph+E%2E+Gonzalez%2E+2025%2E+The+berkeley+function+calling+leaderboard+%28BFCL%29%3A+From+tool+use+to+agentic+eval)
24. Yuhang Liu, Pengxiang Li, Congkai Xie, Xavier Hu, Xiaotian Han, Shengyu Zhang, Hongxia Yang, and Fei Wu. 2025b. Infigui-r1: Advancing multimodal gui agents from reactive actors to deliberative reasoners. arXiv preprint arXiv:2504.14239. — [arXiv:2504.14239](https://arxiv.org/abs/2504.14239) [find](https://scholar.google.com/scholar?q=Yuhang+Liu%2C+Pengxiang+Li%2C+Congkai+Xie%2C+Xavier+Hu%2C+Xiaotian+Han%2C+Shengyu+Zhang%2C+Hongxia+Yang%2C+and+Fei+Wu%2E+2025b%2E+Infigui%2Dr1%3A+Advancing+multimodal+gui+agents+from+reactive+actors+to+deliberative+reasone)
25. Jonas Pfeiffer, Aishwarya Kamath, Andreas Rücklé, Kyunghyun Cho, and Iryna Gurevych. 2021. [Adapterfusion: Non-destructive task composition for transfer learning](https://md.research.elamur.ai/free/en/papers/pfeiffer-2020-adapterfusion-non-destructive-task-composition/). In Proceedings of the 16th conference of the European chapter of the association for computational linguistics: main volume, pages 487–503…
26. Zuxin Liu, Thai Hoang, Jianguo Zhang, Ming Zhu, Tian Lan, Shirley Kokane, Juntao Tan, Weiran Yao, Zhiwei Liu, Yihao Feng, Rithesh Murthy, Liangwei Yang, Silvio Savarese, Juan Carlos Niebles, Huan Wang, Shelby Heinecke, and Caiming Xiong. 2024. Apigen: Automated pipeline for generating verifiable and… — [arXiv:2406.18518](https://arxiv.org/abs/2406.18518) [find](https://scholar.google.com/scholar?q=Zuxin+Liu%2C+Thai+Hoang%2C+Jianguo+Zhang%2C+Ming+Zhu%2C+Tian+Lan%2C+Shirley+Kokane%2C+Juntao+Tan%2C+Weiran+Yao%2C+Zhiwei+Liu%2C+Yihao+Feng%2C+Rithesh+Murthy%2C+Liangwei+Yang%2C+Silvio+Savarese%2C+Juan+Carlos+Niebles%2C+Huan+Wang)
27. Quanfeng Lu, Wenqi Shao, Zitao Liu, Lingxiao Du, Fanqing Meng, Boxuan Li, Botong Chen, Siyuan Huang, Kaipeng Zhang, and Ping Luo. 2025a. Guiodyssey: A comprehensive dataset for cross-app gui navigation on mobile devices. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pag… [find](https://scholar.google.com/scholar?q=Quanfeng+Lu%2C+Wenqi+Shao%2C+Zitao+Liu%2C+Lingxiao+Du%2C+Fanqing+Meng%2C+Boxuan+Li%2C+Botong+Chen%2C+Siyuan+Huang%2C+Kaipeng+Zhang%2C+and+Ping+Luo%2E+2025a%2E+Guiodyssey%3A+A+comprehensive+dataset+for+cross%2Dapp+gui+navigatio)
28. Zhengxi Lu, Yuxiang Chai, Yaxuan Guo, Xi Yin, Liang Liu, Hao Wang, Han Xiao, Shuai Ren, Guanjing Xiong, and Hongsheng Li. 2025b. Uir1: Enhancing efficient action prediction of gui agents by reinforcement learning. arXiv preprint arXiv:2503.21620. — [arXiv:2503.21620](https://arxiv.org/abs/2503.21620) [find](https://scholar.google.com/scholar?q=Zhengxi+Lu%2C+Yuxiang+Chai%2C+Yaxuan+Guo%2C+Xi+Yin%2C+Liang+Liu%2C+Hao+Wang%2C+Han+Xiao%2C+Shuai+Ren%2C+Guanjing+Xiong%2C+and+Hongsheng+Li%2E+2025b%2E+Uir1%3A+Enhancing+efficient+action+prediction+of+gui+agents+by+reinforcem)
29. Run Luo, Lu Wang, Wanwei He, Longze Chen, Jiaming Li, and Xiaobo Xia. 2025. Gui-r1: A generalist r1-style vision-language action model for gui agents. arXiv preprint arXiv:2504.10458. — [arXiv:2504.10458](https://arxiv.org/abs/2504.10458) [find](https://scholar.google.com/scholar?q=Run+Luo%2C+Lu+Wang%2C+Wanwei+He%2C+Longze+Chen%2C+Jiaming+Li%2C+and+Xiaobo+Xia%2E+2025%2E+Gui%2Dr1%3A+A+generalist+r1%2Dstyle+vision%2Dlanguage+action+model+for+gui+agents%2E+arXiv+preprint+arXiv%3A2504%2E10458%2E+%E2%80%94+%5BarXiv%3A2504%2E10)
30. Xinbei Ma, Zhuosheng Zhang, and Hai Zhao. 2024. Coco-agent: A comprehensive cognitive mllm agent for smartphone gui automation. In Findings of the Association for Computational Linguistics ACL 2024, pages 9097–9110. [find](https://scholar.google.com/scholar?q=Xinbei+Ma%2C+Zhuosheng+Zhang%2C+and+Hai+Zhao%2E+2024%2E+Coco%2Dagent%3A+A+comprehensive+cognitive+mllm+agent+for+smartphone+gui+automation%2E+In+Findings+of+the+Association+for+Computational+Linguistics+ACL+2024%2C+p)
31. Seyed Iman Mirzadeh, Mehrdad Farajtabar, Dilan Gorur, Razvan Pascanu, and Hassan Ghasemzadeh. [find](https://scholar.google.com/scholar?q=Seyed+Iman+Mirzadeh%2C+Mehrdad+Farajtabar%2C+Dilan+Gorur%2C+Razvan+Pascanu%2C+and+Hassan+Ghasemzadeh%2E)
32. Aske Plaat, Annie Wong, Suzan Verberne, Joost Broekens, Niki van Stein, and Thomas Back. 2024. [Reasoning with large language models, a survey](https://md.research.elamur.ai/free/en/papers/plaat-2024-multi-step-reasoning-large-language/). arXiv preprint arXiv:2407.11511.
33. Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, Sihan Zhao, Lauren Hong, Runchu Tian, Ruobing Xie, Jie Zhou, Mark Gerstein, Dahai Li, Zhiyuan Liu, and Maosong Sun. 2023. [Toolllm: Facilitating large language models to master 16000+ real…](https://md.research.elamur.ai/free/en/papers/qin-2023-toolllm-facilitating-large-language-models/)
34. Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. 2017. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001–2010. [find](https://scholar.google.com/scholar?q=Sylvestre%2DAlvise+Rebuffi%2C+Alexander+Kolesnikov%2C+Georg+Sperl%2C+and+Christoph+H+Lampert%2E+2017%2E+icarl%3A+Incremental+classifier+and+representation+learning%2E+In+Proceedings+of+the+IEEE+conference+on+Computer)
35. Anthony Robins. 1995. Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 7(2):123–146. [find](https://scholar.google.com/scholar?q=Anthony+Robins%2E+1995%2E+Catastrophic+forgetting%2C+rehearsal+and+pseudorehearsal%2E+Connection+Science%2C+7%282%29%3A123%E2%80%93146%2E)
36. Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, 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/). Preprint, arXiv:2302.04761.
37. Jonathan Schwarz, Wojciech Czarnecki, Jelena Luketina, Agnieszka Grabska-Barwinska, Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. 2018. [Progress & compress: A scalable framework for continual learning](https://md.research.elamur.ai/free/en/papers/schwarz-2018-progress-compress-scalable-framework-continual/). In International conference on machine learning, pages 4528–4537. PMLR.
38. Jingwei Shi, Xinxiang Yin, Jing Huang, Jinman Zhao, and Shengyu Tao. 2026. [Codehacker: Automated test case generation for detecting vulnerabilities in competitive programming solutions](https://md.research.elamur.ai/free/en/papers/shi-2026-codehacker-automated-test-case-generation/). arXiv preprint arXiv:2602.20213.
39. Carmon, Simon Kornblith, and 1 others. 2022. [Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time](https://md.research.elamur.ai/free/en/papers/wortsman-2022-model-soups-averaging-weights-multiple/). In International conference on machine learning, pages 23965–23998. PMLR.
40. Chuanneng Sun, Songjun Huang, and Dario Pompili. 2025. Llm-based multi-agent decision-making: Challenges and future directions. IEEE Robotics and Automation Letters. [find](https://scholar.google.com/scholar?q=Chuanneng+Sun%2C+Songjun+Huang%2C+and+Dario+Pompili%2E+2025%2E+Llm%2Dbased+multi%2Dagent+decision%2Dmaking%3A+Challenges+and+future+directions%2E+IEEE+Robotics+and+Automation+Letters%2E)
41. Zhiyong Wu, Zhenyu Wu, Fangzhi Xu, Yian Wang, Qiushi Sun, Chengyou Jia, Kanzhi Cheng, Zichen Ding, Liheng Chen, Paul Pu Liang, and 1 others. 2025. Os-atlas: Foundation action model for generalist gui agents. In The Thirteenth International Conference on Learning Representations. [find](https://scholar.google.com/scholar?q=Zhiyong+Wu%2C+Zhenyu+Wu%2C+Fangzhi+Xu%2C+Yian+Wang%2C+Qiushi+Sun%2C+Chengyou+Jia%2C+Kanzhi+Cheng%2C+Zichen+Ding%2C+Liheng+Chen%2C+Paul+Pu+Liang%2C+and+1+others%2E+2025%2E+Os%2Datlas%3A+Foundation+action+model+for+generalist+gui+)
42. Fei Tang, Zhangxuan Gu, Zhengxi Lu, Xuyang Liu, Shuheng Shen, Changhua Meng, Wen Wang, Wenqi Zhang, Yongliang Shen, Weiming Lu, and 1 others. 2025a. Gui-g2: Gaussian reward modeling for gui grounding. arXiv preprint arXiv:2507.15846. — [arXiv:2507.15846](https://arxiv.org/abs/2507.15846) [find](https://scholar.google.com/scholar?q=Fei+Tang%2C+Zhangxuan+Gu%2C+Zhengxi+Lu%2C+Xuyang+Liu%2C+Shuheng+Shen%2C+Changhua+Meng%2C+Wen+Wang%2C+Wenqi+Zhang%2C+Yongliang+Shen%2C+Weiming+Lu%2C+and+1+others%2E+2025a%2E+Gui%2Dg2%3A+Gaussian+reward+modeling+for+gui+grounding%2E)
43. Fei Tang, Haolei Xu, Hang Zhang, Siqi Chen, Xingyu Wu, Yongliang Shen, Wenqi Zhang, Guiyang Hou, Zeqi Tan, Yuchen Yan, and 1 others. 2025b. A survey on (m) llm-based gui agents. arXiv preprint arXiv:2504.13865. — [arXiv:2504.13865](https://arxiv.org/abs/2504.13865) [find](https://scholar.google.com/scholar?q=Fei+Tang%2C+Haolei+Xu%2C+Hang+Zhang%2C+Siqi+Chen%2C+Xingyu+Wu%2C+Yongliang+Shen%2C+Wenqi+Zhang%2C+Guiyang+Hou%2C+Zeqi+Tan%2C+Yuchen+Yan%2C+and+1+others%2E+2025b%2E+A+survey+on+%28m%29+llm%2Dbased+gui+agents%2E+arXiv+preprint+arXiv%3A2)
44. Yu-Ming Tang, Yi-Xing Peng, and Wei-Shi Zheng. 2023. When prompt-based incremental learning does not meet strong pretraining. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1706–1716. [find](https://scholar.google.com/scholar?q=Yu%2DMing+Tang%2C+Yi%2DXing+Peng%2C+and+Wei%2DShi+Zheng%2E+2023%2E+When+prompt%2Dbased+incremental+learning+does+not+meet+strong+pretraining%2E+In+Proceedings+of+the+IEEE%2FCVF+International+Conference+on+Computer+Vision)
45. Haoming Wang, Haoyang Zou, Huatong Song, Jiazhan Feng, Junjie Fang, Junting Lu, Longxiang Liu, Qinyu Luo, Shihao Liang, Shijue Huang, and 1 others. 2025a. [Ui-tars-2 technical report: Advancing gui agent with multi-turn reinforcement learning](https://md.research.elamur.ai/free/en/papers/wang-2025-ui-tars-2-technical/). arXiv preprint arXiv:2509.02544.
46. Liyuan Wang, Jingyi Xie, Xingxing Zhang, Mingyi Huang, Hang Su, and Jun Zhu. 2023. Hierarchical decomposition of prompt-based continual learning: Rethinking obscured sub-optimality. Advances in Neural Information Processing Systems, 36:69054– 69076. [find](https://scholar.google.com/scholar?q=Liyuan+Wang%2C+Jingyi+Xie%2C+Xingxing+Zhang%2C+Mingyi+Huang%2C+Hang+Su%2C+and+Jun+Zhu%2E+2023%2E+Hierarchical+decomposition+of+prompt%2Dbased+continual+learning%3A+Rethinking+obscured+sub%2Doptimality%2E+Advances+in+Neural)
47. Taiyi Wang, Zhihao Wu, Jianheng Liu, Jianye HAO, Jun Wang, and Kun Shao. 2025b. Distrl: An asynchronous distributed reinforcement learning framework for on-device control agent. In The Thirteenth International Conference on Learning Representations. [find](https://scholar.google.com/scholar?q=Taiyi+Wang%2C+Zhihao+Wu%2C+Jianheng+Liu%2C+Jianye+HAO%2C+Jun+Wang%2C+and+Kun+Shao%2E+2025b%2E+Distrl%3A+An+asynchronous+distributed+reinforcement+learning+framework+for+on%2Ddevice+control+agent%2E+In+The+Thirteenth+Inte)
48. Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. 2022. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 139–149. [find](https://scholar.google.com/scholar?q=Zifeng+Wang%2C+Zizhao+Zhang%2C+Chen%2DYu+Lee%2C+Han+Zhang%2C+Ruoxi+Sun%2C+Xiaoqi+Ren%2C+Guolong+Su%2C+Vincent+Perot%2C+Jennifer+Dy%2C+and+Tomas+Pfister%2E+2022%2E+Learning+to+prompt+for+continual+learning%2E+In+Proceedings+of+)
49. Hui Wei, Zihao Zhang, Shenghua He, Tian Xia, Shijia Pan, and Fei Liu. 2025. [Plangenllms: A modern survey of llm planning capabilities](https://md.research.elamur.ai/free/en/papers/wei-2025-plangenllms-modern-survey-llm-planning/). arXiv preprint arXiv:2502.11221.
50. Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Wujiang Xu, Zujie Liang, Kai Mei, Hang Gao, Juntao Tan, and Yongfeng Zhang. 2025a. [A-mem: Agentic memory for llm agents](https://md.research.elamur.ai/free/en/papers/xu-2025-mem-agentic-memory-llm-agents/). arXiv preprint arXiv:2502.12110.
51. Yifan Xu, Xiao Liu, Xinghan Liu, Jiaqi Fu, Hanchen Zhang, Bohao Jing, Shudan Zhang, Yuting Wang, Wenyi Zhao, and Yuxiao Dong. 2025b. Mobilerl: Online agentic reinforcement learning for mobile gui agents. arXiv preprint arXiv:2509.18119. — [arXiv:2509.18119](https://arxiv.org/abs/2509.18119) [find](https://scholar.google.com/scholar?q=Yifan+Xu%2C+Xiao+Liu%2C+Xinghan+Liu%2C+Jiaqi+Fu%2C+Hanchen+Zhang%2C+Bohao+Jing%2C+Shudan+Zhang%2C+Yuting+Wang%2C+Wenyi+Zhao%2C+and+Yuxiao+Dong%2E+2025b%2E+Mobilerl%3A+Online+agentic+reinforcement+learning+for+mobile+gui+agen)
52. Shipeng Yan, Jiangwei Xie, and Xuming He. 2021. Der: Dynamically expandable representation for class incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3014–3023. [find](https://scholar.google.com/scholar?q=Shipeng+Yan%2C+Jiangwei+Xie%2C+and+Xuming+He%2E+2021%2E+Der%3A+Dynamically+expandable+representation+for+class+incremental+learning%2E+In+Proceedings+of+the+IEEE%2FCVF+conference+on+computer+vision+and+pattern+reco)
53. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, and 41 others. 2025. [Qwen3 technical report](https://md.research.elamur.ai/free/en/papers/yang-2025-qwen3-technical-report/). arXiv preprint arXiv:2505.09388.
54. Jiabo Ye, Xi Zhang, Haiyang Xu, Haowei Liu, Junyang Wang, Zhaoqing Zhu, Ziwei Zheng, Feiyu Gao, Junjie Cao, Zhengxi Lu, and 1 others. 2025. Mobile-agent-v3: Fundamental agents for gui automation. arXiv preprint arXiv:2508.15144. — [arXiv:2508.15144](https://arxiv.org/abs/2508.15144) [find](https://scholar.google.com/scholar?q=Jiabo+Ye%2C+Xi+Zhang%2C+Haiyang+Xu%2C+Haowei+Liu%2C+Junyang+Wang%2C+Zhaoqing+Zhu%2C+Ziwei+Zheng%2C+Feiyu+Gao%2C+Junjie+Cao%2C+Zhengxi+Lu%2C+and+1+others%2E+2025%2E+Mobile%2Dagent%2Dv3%3A+Fundamental+agents+for+gui+automation%2E+arXi)
55. Friedemann Zenke, Ben Poole, and Surya Ganguli. 2017. Continual learning through synaptic intelligence. In International conference on machine learning, pages 3987–3995. PMLR. [find](https://scholar.google.com/scholar?q=Friedemann+Zenke%2C+Ben+Poole%2C+and+Surya+Ganguli%2E+2017%2E+Continual+learning+through+synaptic+intelligence%2E+In+International+conference+on+machine+learning%2C+pages+3987%E2%80%933995%2E+PMLR%2E)
56. Chaoyun Zhang, Shilin He, Jiaxu Qian, Bowen Li, Liqun Li, Si Qin, Yu Kang, Minghua Ma, Guyue Liu, Qingwei Lin, and 1 others. 2024a. Large language model-brained gui agents: A survey. arXiv preprint arXiv:2411.18279. — [arXiv:2411.18279](https://arxiv.org/abs/2411.18279) [find](https://scholar.google.com/scholar?q=Chaoyun+Zhang%2C+Shilin+He%2C+Jiaxu+Qian%2C+Bowen+Li%2C+Liqun+Li%2C+Si+Qin%2C+Yu+Kang%2C+Minghua+Ma%2C+Guyue+Liu%2C+Qingwei+Lin%2C+and+1+others%2E+2024a%2E+Large+language+model%2Dbrained+gui+agents%3A+A+survey%2E+arXiv+preprint+ar)
57. Guibin Zhang, Hejia Geng, Xiaohang Yu, Zhenfei Yin, Zaibin Zhang, Zelin Tan, Heng Zhou, Zhongzhi Li, Xiangyuan Xue, Yijiang Li, and 1 others. 2025a. The landscape of agentic reinforcement learning for llms: A survey. arXiv preprint arXiv:2509.02547. — [arXiv:2509.02547](https://arxiv.org/abs/2509.02547) [find](https://scholar.google.com/scholar?q=Guibin+Zhang%2C+Hejia+Geng%2C+Xiaohang+Yu%2C+Zhenfei+Yin%2C+Zaibin+Zhang%2C+Zelin+Tan%2C+Heng+Zhou%2C+Zhongzhi+Li%2C+Xiangyuan+Xue%2C+Yijiang+Li%2C+and+1+others%2E+2025a%2E+The+landscape+of+agentic+reinforcement+learning+for)
58. Jiwen Zhang, Jihao Wu, Teng Yihua, Minghui Liao, Nuo Xu, Xiao Xiao, Zhongyu Wei, and Duyu Tang. 2024b. Android in the zoo: Chain-of-action-thought for gui agents. In Findings of the Association for Substituting (9) into (10): Computational Linguistics: EMNLP 2024, pages 12016–12031. [find](https://scholar.google.com/scholar?q=Jiwen+Zhang%2C+Jihao+Wu%2C+Teng+Yihua%2C+Minghui+Liao%2C+Nuo+Xu%2C+Xiao+Xiao%2C+Zhongyu+Wei%2C+and+Duyu+Tang%2E+2024b%2E+Android+in+the+zoo%3A+Chain%2Dof%2Daction%2Dthought+for+gui+agents%2E+In+Findings+of+the+Association+for+Su)
59. Kai Zhang, Xiangchao Chen, Bo Liu, Tianci Xue, Zeyi Liao, Zhihan Liu, Xiyao Wang, Yuting Ning, Zhaorun Chen, Xiaohan Fu, and 1 others. 2025b. Agent learning via early experience. arXiv preprint arXiv:2510.08558. θ new = θ pre − η (12) θ new = θ − η∇L̃(θ). [find](https://scholar.google.com/scholar?q=Kai+Zhang%2C+Xiangchao+Chen%2C+Bo+Liu%2C+Tianci+Xue%2C+Zeyi+Liao%2C+Zhihan+Liu%2C+Xiyao+Wang%2C+Yuting+Ning%2C+Zhaorun+Chen%2C+Xiaohan+Fu%2C+and+1+others%2E+2025b%2E+Agent+learning+via+early+experience%2E+arXiv+preprint+arXiv%3A)
60. (13) θ new = θ + For small ∥δ∥: w k ⊙ τ k . (14) w k ⊙ ∇L k (θ). (15) w k ⊤ L k (θ). (16) From (9) and (14): θ new = θ − η K X k=1 From (13) and (15): L̃(θ) = A.2 K X k=1 Detailed Proof of Theorem 2 [find](https://scholar.google.com/scholar?q=%2813%29+%CE%B8+new+%3D+%CE%B8+%2B+For+small+%E2%88%A5%CE%B4%E2%88%A5%3A+w+k+%E2%8A%99+%CF%84+k+%2E+%2814%29+w+k+%E2%8A%99+%E2%88%87L+k+%28%CE%B8%29%2E+%2815%29+w+k+%E2%8A%A4+L+k+%28%CE%B8%29%2E+%2816%29+From+%289%29+and+%2814%29%3A+%CE%B8+new+%3D+%CE%B8+%E2%88%92+%CE%B7+K+X+k%3D1+From+%2813%29+and+%2815%29%3A+L%CC%83%28%CE%B8%29+%3D+A%2E2+K+X+k%3D1+Detailed+Proof+of+Theorem+2)

*Updated 2 November 2026*