已审阅并标注「鲁棒性」的研究与资产。
条目依据已审阅标签归类，同一主题不表示资源之间存在依赖关系。日期保留来源精度，注释保留原语言。
Defines visual state reliance and probes whether GUI agents resolve screenshot/structure conflicts using pixels, serialized interface structure or prior beliefs. Paired interventions and live multi-step tests connect misattributed state beliefs to failures and compare consistency-based mitigations.
Describes a mobile GUI agent trained through real-device interaction, corrected failure trajectories, supervised learning, and reinforcement learning.
Uses failed computer-use trajectories to diagnose failure modes and generate inference-time agent code patches, with light human verification. An OSWorld study evaluates this failure-driven improvement loop without additional model training.
Argus benchmarks post-hoc uncertainty methods for single-step GUI grounding across open-weight and closed-source observation regimes. It tests calibration, rejection, error severity and conformal click regions, finding that uncertainty rankings transfer less reliably across model/interface regimes than across datasets for one model.
The framework uses a vision-language judge’s final-screen assessment as a terminal reward for GUI reinforcement learning. It models evaluator noise and corrects the reward estimator used by PPO across desktop benchmarks.
This benchmark matches desktop goals, initial states and final-state verifiers across screen-only GUI and skill-mediated CLI agents. Its experiments separate graphical execution failures from missing CLI skill coverage, including a verifier-guided skill-augmentation condition.
ENVS branches over GUI actions in live desktop VMs, verifies successful leaves, and converts them into balanced step-level training supervision. It also introduces OSWorld-Noisy to test recovery from dynamic interruptions while retaining the underlying tasks.
Skill-Guided Continuation Distillation first lets a policy reach its own off-trajectory states, then uses skill guidance to collect successful continuations. Mixing those continuations with expert traces supplies supervision for recovery states absent from demonstrations.
MIRAGE evaluates targeted next-action hijacking from adversarial images restricted to legitimate third-party webpage regions. It combines diffusion-based image generation with execution-aware optimization and tests the constrained visual-injection setting on web-agent frameworks.
FraudSMSWalker tests browser agents on bilingual SMS-to-webpage fraud judgments while hiding URL and reputation shortcuts. Its controlled chains include difficult benign login/payment cases and evaluate whether classifications remain supported by the visible message and page evidence.
Compares online skill and memory augmentations with a vanilla web agent given the same total inference-token budget. Across WebArena and WorkArena settings, it studies when module overhead removes aggregate gains and how rerun variability changes conclusions.
VISTA forms GRPO comparison groups from target-preserving crops of the same GUI instance with exactly remapped target boxes. A conditional, advantage-weighted cross-view anchor supplies additional supervision without entering the group baseline, and evaluation includes view-consistency robustness.
This study finds that full-image visual memory can reduce some state errors while worsening action-level failures. Its AGMem alternative stores local crops associated with successful actions or recovery and evaluates their use in desktop GUI tasks.
Speculative Rollback Correction lets a student web agent execute short branches before teacher review localizes harmful deviations. Verified continuations and localized corrections feed supervised training while an archive retains diverse successful solutions.
Workflow-GYM evaluates long GUI workflows in professional software rather than isolated interface actions. Its analysis examines failures such as omitted stages, accumulated errors and goal drift during specialized end-to-end work.
GUI-AC uses grounding certainty to adjust advantage weighting and policy-clipping bounds during continual reinforcement fine-tuning. Its mechanisms aim to suppress noisy confidence updates while retaining exploration under shifting GUI domains and resolutions.
MacArena runs computer-use tasks through Apple’s virtualization framework on Apple Silicon, combining task ports with macOS-native workloads. It distinguishes the ported and native subsets when examining agent performance and cross-platform transfer.
Introduces CUA-HandCrafted, a browser-agent red-teaming benchmark that varies attack templates, tasks and system prompts. Its domain comparison and reproducibility audit caution against transferring browser-surface safety conclusions to coding agents or unreleased optimized injections.
GUI-SAGE generates reasoning conditioned on known actions and weights policy updates using confidence and reward signals. Its experiments use AndroidControl and GUIOdyssey data and separately evaluate interactive AndroidWorld tasks.
MIRAGE constructs mobile screenshot injections in ordinary user-content regions through localization, contextual payload generation and realism curation. Its evaluation separates visual realism from attack effectiveness against pixel-observing GUI agents.
AndroidDaily evaluates mobile GUI agents on daily workflows in closed-source Android apps. Its GRADE evaluator follows visual trajectories against operational obligations, output-quality requirements and negative constraints, producing process-level diagnostics without hidden app state.
MobileExplorer probes relevant UI elements while an on-device model is reasoning, then supplies the exploration trace as compact context. A two-level rollback mechanism restores the working state when simple backtracking fails.
Claw-Anything benchmarks persistent personal assistance across activity histories, backend services and integrated GUI/CLI interaction. Its simulated event histories introduce irrelevant or conflicting context, and a separate generation pipeline supplies training environments.
Weasel selects web-agent training steps by balancing importance and diversity across states, sites and interaction patterns. It also prunes training accessibility trees around known action targets and regenerates rationales to match the model’s reasoning style.
WinDeskGround synthesizes multi-window desktop layouts with controlled occlusion, density and semantic similarity to test GUI grounding robustness. Its instruction-target benchmark measures how localization changes when interfaces depart from simple single-layer scenes.
DUDE combines asymmetric hybrid-reward learning with summaries of deceptive-interface failures to guide safer web actions. Its RUC benchmark evaluates deception susceptibility alongside ordinary task performance across interface scenarios.
Introduces Faithful-Agent, which first teaches abstention under perturbed evidence and then refines GUI policies with an anchor-based, variance-adaptive advantage estimator. A thought-action consistency reward encourages decisions grounded in displayed evidence and instructions.
Proposes a step-level model cascade that runs a smaller computer-use policy until learned monitors flag stalled progress or semantic drift. Separate history-based stuck detection and milestone checks determine when to escalate for recovery or verification without retraining the larger model.
Provides an evaluation toolkit for joint safety, efficiency, and multimodal robustness analysis of operating-system GUI agents.
VLAA-GUI combines mandatory checks for visually supported completion, multi-tier loop recovery and on-demand workflow search. Optional coding and grounding agents support execution, with component ablations on Linux and Windows GUI tasks.
This study examines UI changes between a desktop agent’s screenshot observation and action dispatch. Its pre-execution state checks address overlay and focus manipulation, while the reported DOM-injection blind spot limits the defense’s coverage.
This study examines repeated OSWorld executions to separate execution stochasticity, task-specification ambiguity and behavioral variation. Paired task-level tests show why one successful run is insufficient evidence of reliable computer use.
WebUncertainty adapts planning mode to task uncertainty and guides action-level tree search with confidence-derived uncertainty estimates. Its evaluations examine autonomous execution of dynamic, long-horizon web tasks on WebArena and WebVoyager.
DailyDroid compares text-only and text-plus-screenshot agents on everyday Android tasks across multiple apps and difficulty levels. Its failure analysis examines accessibility, observation modality and agent/application behavior.
Introduces HalluClear with a GUI hallucination taxonomy, expert-calibrated evaluation workflow and structured-reasoning post-training scheme. It evaluates how these components affect hallucinations, grounding and action fidelity across GUI agents.
Introduces FedGUI, a federated GUI-learning benchmark built from six datasets spanning mobile, web and desktop interfaces. Its experiments separate platform, device, operating-system and data-source heterogeneity when evaluating collaborative training.
Introduces PriceBlind, a visual perturbation method for controlled screenshot-based evaluation of price-constrained multimodal agents. The reported transfer experiments use a simplified single-turn coordinate-selection protocol.
Introduces GUI-Perturbed, which varies screenshot appearance and instruction wording independently to diagnose grounding failures. Experiments on three related model architectures examine relational instructions, browser zoom and augmentation-based fine-tuning.
Studies iterative cursor grounding in dense code-editor interfaces, using visual feedback from earlier attempts to correct displacement. PrecisionCUA evaluates this closed-loop process across multiple model families on coding-interface tasks.
This study tests black-box GUI-grounding failures caused by strategically added, innocuous-looking UI elements. Its iterative overlay search is compared with random injection and repeated independent grounding trials to distinguish persistent misdirection from incidental clutter.
Introduces VeriGUI, which checks expected action effects and reasons about recovery before subsequent GUI actions. Training combines synthetic failure trajectories with verification-weighted GRPO, evaluated on an AndroidControl-based robustness benchmark.
Presents demonstration-based GUI process replay with Sequential Monte Carlo localization and readiness checks. GPA executes the learned workflow locally and can serve as the GUI execution component of an agent using MCP or CLI tools.
AgentRAE studies notification-based visual backdoors that redirect mobile GUI-agent actions. Its training separates sensitivity to icon differences from trigger/action association, with evaluations of clean behavior and targeted redirection.
Turns draft web-agent skills into structured executable artifacts that support deterministic checks, fault localization and local repair. VisualWebArena and MiniWoB experiments examine execution and reuse within the same benchmark family.
This public red-teaming study evaluates indirect prompt injection that both changes agent behavior and hides the compromise from the final response. Its scenarios include computer use alongside tool calling and coding, with analyses of transfer and model-dependent vulnerability.
Formalizes a visual confused-deputy failure in which a computer agent acts on a misperceived UI object. A separate guardrail checks both the visual target and the agent's action reasoning against deployment-specific knowledge before permitting execution.
Introduces HATS, which directs GUI trajectory exploration toward semantically ambiguous actions and iteratively checks instruction-execution alignment. Feedback from trajectory refinement updates the exploration signal used to collect subsequent training examples.
CUAAudit examines vision-language models that judge computer-use task completion from an instruction and final environment state. Across macOS, Windows and Linux benchmarks, it measures accuracy, confidence calibration and disagreement between auditors.
Introduces AndroidWorld-Generalization with separate tests for unseen task instances, templates, and applications. A GRPO training system combines containerized environments with asynchronous trajectory collection to study these generalization regimes.
TimeWarp tests web agents across multiple versions of the same sites, varying visual design, frontend code and workflows. TimeTraj transfers plans across versions to collect training trajectories and study robustness to interface change.
Combines supervised and reinforcement fine-tuning for continual GUI learning, using policy entropy to adjust their proportions and gradient projection to limit conflicting updates. AndroidControl-CL groups applications into sequential tasks to evaluate adaptation and retention.
ACuRL uses autonomous interface exploration and an adaptive task curriculum to continually train computer-use policies with reinforcement learning. CUAJudge supplies outcome feedback, and experiments study adaptation across desktop applications and environment changes.
MisActBench provides action-level labels for task alignment, while DeAction checks proposed computer-use actions and requests revisions through structured feedback. The paper separately evaluates offline detection and interactive operation in OSWorld and RedTeamCUA.
Introduces SafeGround, which estimates GUI-grounding uncertainty from the spatial dispersion of sampled predictions and calibrates a decision threshold before testing. It studies risk-aware selection on ScreenSpot-Pro, with statistical control scoped to the threshold’s false-discovery rate.
BEAP-Agent treats GUI execution as depth-first search with coordinated planning, execution and task-state tracking. It supports backtracking across multiple levels of earlier states and evaluates recovery on OSWorld.
Continual GUI Agents defines grounding under changing application domains and screen resolutions. Its GUI-AiF training framework uses separate point- and region-alignment rewards to adapt grounding as these distributions shift.
Introduces MobileBench-OL, an online benchmark of Chinese-app tasks spanning execution, reasoning, exploration, and noisy conditions. Its automatic evaluation and reset framework supports repeated testing of mobile GUI agents in real environments.
MAGNET maintains separate memories for stable interface functions and task procedures across changing layouts and workflows. It updates frequently used knowledge and evaluates adaptation in AndroidWorld and offline distribution-shift settings.
MagicGUI-RMS combines domain-specific and general reward models to assess GUI actions and propose corrections. A structured data pipeline and feedback-recycling process support repeated reward-model and agent improvement.
MirrorGuard trains a reasoning-correction defense using GUI risk trajectories generated in a text-based neural-symbolic simulator. It evaluates whether corrections transfer to computer-use agents operating real interfaces, keeping simulated training separate from deployment tests.
第 61–120 条，共 186 条资源。
通过主题导出清单获取完整记录和证据。克隆目录后，可将主题记录直接交给标准 RAG 与引用导出工具。
