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。
NOVA separates trusted task planning from screen-derived observations by producing a branching plan before execution and resolving runtime values through perception. Its OSWorld study examines security and utility while identifying remaining risks from perception-driven branch steering.
Studies deceptive social-engineering web patterns targeting browser automation agents and proposes defenses against bait-based manipulation.
AndroidLens provides 571 extended mobile tasks in Chinese and English, with multiple valid paths and preserved environmental anomalies. Its dynamic evaluation uses nested milestones and Average Task Progress to measure partial completion of long workflows.
VenusBench-GD is a bilingual, cross-platform GUI grounding benchmark organized into basic and advanced tasks. Its annotation pipeline and six-subtask taxonomy support comparisons of general multimodal models and GUI-specialized models across different grounding demands.
ceLLMate constrains browser-agent authority through an extension that applies policies at the HTTP layer. It evaluates this sandbox against prompt-injection attacks in WASP, avoiding dependence on brittle low-level click and keystroke policies.
MVP addresses unstable GUI coordinates by predicting across several attention-guided cropped views. It clusters the resulting coordinates and selects the densest cluster centroid, evaluating this training-free inference method across grounding models.
Chain-of-Ground uses iterative visual reasoning and reference feedback to refine GUI localization without additional training. It evaluates on ScreenSpot-Pro and introduces TPanel-UI to probe grounding on industrial control panels with visual distortions.
MPR-GUI evaluates GUI perception and reasoning through strictly aligned tasks in six languages. GUI-XLI adjusts non-English hidden representations toward English representations at language-sensitive layers during inference to address the measured cross-language gaps.
Introduces OpenApps, a configurable app ecosystem for measuring UI-agent reliability under changes in appearance and content. Its evaluation compares task success and failure behavior across many versions of the same apps.
UI-CUBE evaluates 226 tasks ranging from individual controls to copy-paste and enterprise workflows. Interface variations, multiple resolutions, and application-state validation support diagnosis of operational reliability beyond task accuracy alone.
D-GARA evaluates Android GUI agents under runtime anomalies such as permission dialogs, battery warnings, and update prompts. Its modular environment embeds these interruptions in Android applications to study recovery and robustness.
HyperClick trains GUI grounders to produce a click and an accompanying confidence estimate using rewards for correctness and confidence alignment. Its evaluation examines localization, calibration, and abstention behavior.
Surfer2 combines hierarchical context management with separate planning and execution, self-verification, and recovery for visual computer use. It evaluates the framework across web, desktop, and Android tasks without task-specific fine-tuning.
Introduces SecureWebArena with simulated interactive websites, an attack-vector taxonomy and labeled agent trajectories. Its evaluation separates failures in reasoning, behavioral traces and task outcomes across vision-language agent families.
Uses the near-zero contact-touch signals of automated mobile agents to expose malicious visual content selectively during agent interactions. The paper combines this agent-only perceptual injection with HG-IDA* prompt optimization to study unauthorized cross-app actions.
This work addresses GUI localization errors caused by mismatches between visual patch positions and output pixel coordinates. RULER tokens provide explicit coordinate references, while Interleaved MRoPE balances width and height encoding; evaluation uses the ScreenSpot benchmark family.
BLIND-ACT tests whether computer-use agents persist with ambiguous, contradictory, or infeasible goals without adequate contextual reasoning. Its OSWorld-based tasks and judged trajectories examine these behaviors and the effect of mitigation prompts.
Introduces WAInjectBench with a threat-model taxonomy and malicious and benign text-image datasets for web-agent prompt-injection detection. It compares detector families across explicit and less visible forms of injected content.
WAREX augments existing web-agent benchmarks with reliability challenges arising from unstable systems and changing or disruptive website behavior. Evaluation on WebArena, WebVoyager and REAL examines how these disturbances affect task completion.
Evaluates multimodal LLMs on web interface reasoning, visual perturbation robustness, and safety-critical web understanding tasks.
GUI-Cursor treats GUI grounding as a sequence of cursor movements guided by visible feedback and movement history. Online reinforcement learning with trajectory rewards trains the model to refine the target location across multiple steps.
Presents an agent-engineering framework combining runtime goal clarification, an observable component architecture and evolving execution memory. The Fairy mobile GUI agent evaluates these mechanisms on ambiguous tasks and through controlled and user studies.
Introduces Chameleon, an environmental-injection evaluation framework that varies webpage content and trigger context. It combines simulated page variation with attention-based optimization and tests GUI agents across realistic websites.
NaviMaster trains a shared policy for GUI and embodied navigation using a common visual-trajectory formulation. It combines trajectory collection, reinforcement learning and distance-aware rewards, with evaluation across both task families.
MobileUse reflects on errors at action and task timescales, invoking reflection when needed and using proactive exploration to learn unfamiliar apps. It evaluates this mobile execution framework on AndroidWorld and AndroidLab and describes a toolkit for physical-device operation.
Introduces a finer-grained taxonomy of GUI localization errors and a logit-based Peak Sharpness Score for analyzing coordinate-prediction uncertainty. A context-aware cropping method refines the visual input without retraining the grounding model.
Builds GUI task data with real-world interaction anomalies and evaluates how agent behavior changes under those conditions.
Demonstrates clean-text visual backdoor attacks (VIBMA) that compromise fine-tuned VLM mobile agents through visual triggers while keeping text instructions intact.
GUI-Reflection synthesizes reflection and error-correction examples from successful trajectories and adds dedicated learning stages for these behaviors. It introduces a reflection task suite, a mobile environment for online data collection, and iterative reflection tuning.
Introduces GUIPilot, which compares mobile design mock-ups with implemented screens through widget alignment and tests expected transitions by executing inferred UI actions. It evaluates both visual inconsistencies and workflow-specific bugs.
Uses a pre-execution critic and reinforcement training to evaluate proposed GUI actions and provide corrective suggestions before execution.
Benchmarks truthfulness, controllability, safety, and privacy of multimodal LLM agents across interactive web and mobile GUI tasks.
Introduces UI-Evol, which extracts action sequences from agent interactions and critiques external guidance against those traces. Evaluation on OSWorld examines how revised operational knowledge affects execution and behavioral consistency.
BacktrackAgent adds verification, judgment, and reflection modules to detect erroneous GUI actions and guide recovery. It develops backtracking-oriented training data that includes post-action pages and evaluates task completion and step accuracy on Mobile3M and Auto-UI.
GEM detects instructions outside a GUI agent’s capabilities by modeling distances in its input-embedding space with a Gaussian mixture. Experiments across mobile, computer and browser data also examine selective escalation to a cloud model when an input is detected as unfamiliar.
Mobile-Bench-v2 evaluates mobile GUI agents using multiple acceptable offline action paths, noisy interfaces with ads or pop-ups, and ambiguous instructions with scripted clarification exchanges. These splits separate step-level execution, robustness, and proactive interaction.
WebRollback gives a web agent an explicit action for returning to an earlier state in its navigation trajectory. It studies this recovery mechanism on two live web-navigation benchmarks under both zero-shot and fine-tuned settings.
Evaluates GUI grounding models under natural image noise, untargeted adversarial perturbations, and targeted image attacks. Experiments cover mobile, desktop, and web interfaces and analyze sensitivity to image resolution and adversarial changes.
Examines active environmental injection against mobile agents, testing whether adversarial interface content redirects decisions during task execution.
Introduces desktop and web tasks with varied initial states to test planning recovery and adaptability beyond canonical starting configurations.
Builds self-critique training examples from searched recovery paths to improve agents' ability to detect and correct errors during interaction.
OSCAR translates user instructions into Python control code while observing screen images and acting through mouse and keyboard operations. A state-machine controller handles errors and replans tasks during execution across desktop and mobile applications.
Studies text and visual validation for web agents and uses the resulting validator to detect failures and guide workflow self-refinement.
Dynamic Planning of Thoughts revises a mobile GUI plan using current environmental feedback and execution history. The study compares this planning strategy with long-dialogue prompting and examines its behavior on unfamiliar tasks and different language-model backbones.
Measures how irrelevant GUI content affects multimodal agents using a simulated dataset and three perception patterns. The study also examines adversarial environment injection and possible ways to improve instruction faithfulness.
Introduces adversarial tasks on VisualWebArena and an evaluation framework that tracks how manipulated information passes between agent components. The analysis examines vulnerabilities introduced by reflection and tree-search evaluators as well as direct action generation.
Introduces B-MoCA, an interactive mobile-control benchmark that varies device settings, layouts, and languages to test generalization.
WILBUR learns to rank prior task demonstrations for a web agent’s prompt and synthesizes instructions from past runs. A generated task curriculum and a backtracking mechanism support adaptation and recovery, evaluated on WebVoyager.
Benchmarks robustness of API-based PowerPoint completion under sentence-level, semantic, and multilingual instruction perturbations and changes to the available API set.
Presents a conceptual perception–reasoning–action framework and 12 potential attack scenarios for language agents, including hypothetical web-navigation attacks.
CompWoB adds 50 composed web-automation tasks to study generalization from simpler interactions. The paper compares prompted and fine-tuned agents, balances task training data, and probes sensitivity to changes in instruction order.
LASER represents web navigation as exploration over a defined set of states, allowing an agent to revisit earlier states after mistakes. The study evaluates this backtracking approach on WebShop and Amazon tasks.
Introduces Ringer, which converts demonstrated webpage interactions into scripts combining actions, target selectors, and waiting conditions. It evaluates replay robustness on tasks recorded from real websites.
Evaluation scripts and linked data for comparing GUI action predictions and task execution under ordinary and anomalous interface conditions.
Measures hallucination in UI-element parsing by matching the boxes and descriptions a model lists for 1,000 English and 1,000 Chinese screenshots against human annotations.
Unified evaluation framework measuring truthfulness, controllability, safety, and privacy of multimodal LLM agents across interactive web and mobile GUI tasks.
Offline mobile-agent task data spanning common tasks and released noisy, contaminated and ambiguous-instruction subsets.
Configurable browser apps and task rewards for testing computer-use agents under changes in content, layout and appearance.
第 121–180 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。
