Tag
Microsoft Mage Team presents Mage-VL-Cowork, a full training recipe (data synthesis from evolved harnesses, multi-stage SFT, and online hybrid RL) for learning hybrid computer use, with the 4B model achieving strong results on OSWorld-Verified relative to its size.
提出 SolveBind 框架，将 CAPTCHA 验证分解为层级求解树并配合运行时监督器，在九类 CAPTCHA 任务与多个 GUI 智能体骨干模型上显著提升端到端成功率，同时揭示了恢复机制的成功率与成本权衡。
本文提出 TAG-Nav 框架，通过任务无关的图导航记忆为移动 GUI 代理提供可复用的导航知识，并引入 DestBench 导航评测协议，实验表明图记忆可显著提升 UI-TARS-1.5-7B 和 Qwen-VL-Max 的导航成功率。
ECHOES 提出一种通过递归历史压缩与特权在线蒸馏来增强长时程 GUI 智能体状态保持能力的方法，无需在部署时保留历史截图即可取得有竞争力的性能，并发布了细粒度诊断基准 ECHOESBMK。
The paper introduces FocusGUI, a lightweight adapter that dynamically selects relevant multimodal evidence for GUI agents to prevent compounding errors from irrelevant context, improving action accuracy without fine-tuning the backbone.
阿里巴巴 MAI-UI 团队提出 MAI-Gym，一个开源的一体化平台，统一 GUI 智能体的评估、数据收集与在线强化学习训练，覆盖移动、电脑和 Web 环境，并训练出在 MobileWorld 和 OSWorld-Verified 上达到 SOTA 的 9B 模型 MAI-Gym-9B。
该论文系统研究了VLM驱动的GUI智能体中的记忆模态选择问题，发现最优记忆表示是随交互步骤变化的，并提出SwitchMem路由框架在每一步动态选择文本或视觉记忆，在三个GUI基准上超越了使用静态记忆表示的基线方法。
Work2Skill 提出一个框架，通过 WorkRecorder 收集人类 GUI 操作演示，构建 WorkDemo-350 多模态数据集，并用智能体式技能归纳方法从人类经验中提炼可复用的智能体技能，在 MyPCBench 和 OSWorld-Verified 上显著提升了多种执行模型的任务表现。
A brief X post joking that computer use appears to be "solved," quoting a comparison showing that large general models point on computer UIs better than specialized models due to richer vocabulary.
MemoWM 提出一种以世界模型为条件的记忆分配框架，通过共享预测压缩智能体长期经验存储，在五个长时程智能体记忆基准上达到42.42%平均准确率，同时将经验特定存储降低53.9%。研究还揭示了共享模型容量与重复存储成本之间的权衡关系。
该论文提出将 GUI 世界模型的预测目标从全屏像素重建转移到潜空间中的动作诱导未来信息，通过共享的多模态模型表示配合动作头与转移头，并引入动作感知对比目标来缓解身份捷径问题，相比像素空间世界模型最高实现 31.9% 和 40.6% 的成功率相对提升，并将推理延迟降低最多 98.9%。
The paper introduces LGWM, a decoder-free, action-conditioned world model that predicts next-screen representations in latent space to verify GUI transitions and detect hijacked screens, achieving 0.987 AUC at 17ms per decision on their RSWT-BENCH benchmark, outperforming generative GUI world models with vastly lower latency.
This paper proposes action-conditioned bisimulation over an empirical predictive state graph to decide when GUI agent memories of two web pages should be merged, improving success on MiniWoB++ over memoryless baselines without any training.
The paper introduces AutoGUIWorld, a data-generation framework that uses image generators as visual world models to synthesize GUI interaction trajectories without running software, yielding 79,266 annotated samples that fine-tuning Qwen3.5-35B-A3B to improve OSWorld task score from 33.0% to 40.8% and ScienceBoard success rate from 14.0% to 32.2%.
提出了 ExpActivator，一个免训练框架，通过在冻结的 GUI 骨干网络的潜在空间中匹配当前屏幕与历史状态，仅激活与当前情境相关的经验，而非盲目追加任务相关历史；在四个 GUI 骨干网络上将轨迹内步骤成功率平均提升 28%，同时仅使用约五分之一的历史 token。
The paper introduces GUI-SD-v2, a two-stage on-policy self-distillation framework for multi-turn GUI agents that enhances privilege following and guidance, achieving superior performance on AndroidWorld and MobileWorld benchmarks.
OSWorld-Pro introduces a process-based benchmark with over 300 tasks and 2,800 subgoals, grounded in 67,000+ human annotations, to evaluate computer-use agents step-by-step rather than only by end-state outcomes. It reveals failure modes even in state-of-the-art models, with top performers like Claude Opus 5 scoring 75.7% versus 83.4% on OSWorld.
This paper empirically investigates the susceptibility of LLM-based GUI agents to digital nudges, finding that reasoning configuration redirects rather than reduces nudge effects, positioning interface design as a governance concern for autonomous AI.
This paper introduces EvoSkill-GUI, a training-free framework that allows GUI agents to improve skills through in-execution reflection, revision, and reuse, demonstrating performance gains on multiple benchmarks without retraining.
EchoPath introduces a model-agnostic memory system for GUI agents that replays validated execution trajectories, significantly reducing token cost and execution time for enterprise recurrent tasks.