@inproceedings{li-etal-2026-fully,
title = "Fully Decoupled Reinforcement Learning with Adaptive Data Curation for {GUI} Agents",
author = "Li, Pengxiang and
Zechenhu and
Shang, Zirui and
Wu, Jingrong and
Liu, Yang and
Liu, Hui and
Gao, Zhi and
Shi, Chenrui and
Zhang, Bofei and
Zhang, Zihao and
Shi, Xiaochuan and
YU, Zedong and
Wu, Yuwei and
Wu, Xinxiao and
Jia, Yunde and
Xiang, Liuyu and
He, Zhaofeng and
Li, Qing",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.1654/",
pages = "37604--37633",
ISBN = "979-8-89176-459-0",
abstract = "Vision-Language Models (VLMs) have catalyzed the development of Graphical User Interface (GUI) agents for desktop automation. However, reinforcement learning (RL) for GUI agents remains bottlenecked by two challenges: existing RL frameworks still leave substantial idle time inside the rollout stage, and the learning signal from long-horizon, sparse-reward interactions is highly imbalanced. In this paper, we introduce DART, a fully Decoupled Agentic RL Training framework for GUI agents. Whereas prior RL designs only separate the trainer from the rollout, DART splits the pipeline into four independent modules: environment cluster, rollout service, data manager, and trainer, which brings two extra efficiency benefits: rollout-wise trajectory scheduling and per-worker model synchronization. This yields a $1.6\times$ increase in GPU utilization, a $1.9\times$ boost in training throughput, and a $5.5\times$ increase in environment utilization over a non-decoupled baseline. We further propose a multi-level adaptive data curation scheme operating at the task, trajectory, step, and token levels: it reallocates rollout budget by task difficulty, bootstraps hard tasks with a pre-collected experience pool, restricts policy updates to high-entropy critical steps, and applies truncated importance sampling to absorb the distribution gap between th rollout and training engines. On the OSWorld benchmark, our DART-GUI-7B achieves a 42.13 task success rate, a 14.61 absolute improvement over its base model and a new state of the art among open-source 7B-scale agents. Code, data, and model checkpoints are open-sourced at https://dart-gui-anonymous.github.io (anonymous)."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="li-etal-2026-fully">
<titleInfo>
<title>Fully Decoupled Reinforcement Learning with Adaptive Data Curation for GUI Agents</title>
</titleInfo>
<name type="personal">
<namePart type="given">Pengxiang</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name>
<namePart>Zechenhu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zirui</namePart>
<namePart type="family">Shang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jingrong</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yang</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hui</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhi</namePart>
<namePart type="family">Gao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chenrui</namePart>
<namePart type="family">Shi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Bofei</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zihao</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaochuan</namePart>
<namePart type="family">Shi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zedong</namePart>
<namePart type="family">YU</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yuwei</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xinxiao</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yunde</namePart>
<namePart type="family">Jia</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Liuyu</namePart>
<namePart type="family">Xiang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhaofeng</namePart>
<namePart type="family">He</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Qing</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>Vision-Language Models (VLMs) have catalyzed the development of Graphical User Interface (GUI) agents for desktop automation. However, reinforcement learning (RL) for GUI agents remains bottlenecked by two challenges: existing RL frameworks still leave substantial idle time inside the rollout stage, and the learning signal from long-horizon, sparse-reward interactions is highly imbalanced. In this paper, we introduce DART, a fully Decoupled Agentic RL Training framework for GUI agents. Whereas prior RL designs only separate the trainer from the rollout, DART splits the pipeline into four independent modules: environment cluster, rollout service, data manager, and trainer, which brings two extra efficiency benefits: rollout-wise trajectory scheduling and per-worker model synchronization. This yields a 1.6\times increase in GPU utilization, a 1.9\times boost in training throughput, and a 5.5\times increase in environment utilization over a non-decoupled baseline. We further propose a multi-level adaptive data curation scheme operating at the task, trajectory, step, and token levels: it reallocates rollout budget by task difficulty, bootstraps hard tasks with a pre-collected experience pool, restricts policy updates to high-entropy critical steps, and applies truncated importance sampling to absorb the distribution gap between th rollout and training engines. On the OSWorld benchmark, our DART-GUI-7B achieves a 42.13 task success rate, a 14.61 absolute improvement over its base model and a new state of the art among open-source 7B-scale agents. Code, data, and model checkpoints are open-sourced at https://dart-gui-anonymous.github.io (anonymous).</abstract>
<identifier type="citekey">li-etal-2026-fully</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.1654/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>37604</start>
<end>37633</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Fully Decoupled Reinforcement Learning with Adaptive Data Curation for GUI Agents
%A Li, Pengxiang
%A Shang, Zirui
%A Wu, Jingrong
%A Liu, Yang
%A Liu, Hui
%A Gao, Zhi
%A Shi, Chenrui
%A Zhang, Bofei
%A Zhang, Zihao
%A Shi, Xiaochuan
%A YU, Zedong
%A Wu, Yuwei
%A Wu, Xinxiao
%A Jia, Yunde
%A Xiang, Liuyu
%A He, Zhaofeng
%A Li, Qing
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%A Zechenhu
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F li-etal-2026-fully
%X Vision-Language Models (VLMs) have catalyzed the development of Graphical User Interface (GUI) agents for desktop automation. However, reinforcement learning (RL) for GUI agents remains bottlenecked by two challenges: existing RL frameworks still leave substantial idle time inside the rollout stage, and the learning signal from long-horizon, sparse-reward interactions is highly imbalanced. In this paper, we introduce DART, a fully Decoupled Agentic RL Training framework for GUI agents. Whereas prior RL designs only separate the trainer from the rollout, DART splits the pipeline into four independent modules: environment cluster, rollout service, data manager, and trainer, which brings two extra efficiency benefits: rollout-wise trajectory scheduling and per-worker model synchronization. This yields a 1.6\times increase in GPU utilization, a 1.9\times boost in training throughput, and a 5.5\times increase in environment utilization over a non-decoupled baseline. We further propose a multi-level adaptive data curation scheme operating at the task, trajectory, step, and token levels: it reallocates rollout budget by task difficulty, bootstraps hard tasks with a pre-collected experience pool, restricts policy updates to high-entropy critical steps, and applies truncated importance sampling to absorb the distribution gap between th rollout and training engines. On the OSWorld benchmark, our DART-GUI-7B achieves a 42.13 task success rate, a 14.61 absolute improvement over its base model and a new state of the art among open-source 7B-scale agents. Code, data, and model checkpoints are open-sourced at https://dart-gui-anonymous.github.io (anonymous).
%U https://aclanthology.org/2026.emnlp-main.1654/
%P 37604-37633
Markdown (Informal)
[Fully Decoupled Reinforcement Learning with Adaptive Data Curation for GUI Agents](https://aclanthology.org/2026.emnlp-main.1654/) (Li et al., EMNLP 2026)
ACL
- Pengxiang Li, Zechenhu, Zirui Shang, Jingrong Wu, Yang Liu, Hui Liu, Zhi Gao, Chenrui Shi, Bofei Zhang, Zihao Zhang, Xiaochuan Shi, Zedong YU, Yuwei Wu, Xinxiao Wu, Yunde Jia, Liuyu Xiang, Zhaofeng He, and Qing Li. 2026. Fully Decoupled Reinforcement Learning with Adaptive Data Curation for GUI Agents. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 37604–37633, Budapest, Hungary. Association for Computational Linguistics.