Source-backed papers on GUI grounding, agents, training, evaluation, and safety.
Read the research survey ↗ · Methods, evaluation, and human agency — an English working manuscript with a source-by-source reading register.
AI-assisted editorial review of primary sources; experiments are not independently reproduced. Dates keep their recorded precision. Resource links below follow reviewed introduction or documentation relationships. Method ↗
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.
Introduces CUActSpot to evaluate clicks, drags, drawing and other screen interactions across GUI, text, tables, canvases and natural images. A renderer-based pipeline records geometry and screenshots, then generates matching instructions and action traces for training a grounding model.
LongMemEval-V2 asks memory systems to recover environment knowledge from long web-agent histories for question answering. Its AgentRunbook variants contrast indexed knowledge pools with coding-agent retrieval from stored trajectory files.
ToolCUA synthesizes interleaved GUI/tool trajectories and trains agents to choose between graphical actions and higher-level tools. Staged supervised and reinforcement learning culminate in online hybrid-environment training with a reward for efficient execution paths.
Executable Agentic Memory stores GUI routines in a knowledge graph built through state-aware exploration and action-group mining. A learned value model guides tree search over this executable memory for retrieval-based planning, with AndroidWorld and efficiency evaluation.
Trains GUI agents with a learned selector that removes repeated visual patches across consecutive screenshots while preserving spatial structure. ReVision evaluates this history representation on OSWorld, WebTailBench and AgentNetBench.
Compares mobile world models that predict state changes as delta text, full text, images or renderable code. Downstream Android benchmarks examine synthetic training trajectories and test-time guidance, including limits of post-action reflection.
Re-Prefill adds an attention-guided second prefill pass that revisits candidate screen regions before coordinate generation. The study evaluates this training-free intervention across GUI-grounding models and benchmarks.
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 Mem-W, which compresses previous trajectories and current-session segments into latent memory tokens used directly by the GUI policy. Self-distillation and outcome supervision train this memory representation for web and mobile navigation.
MultiMedia-TerminalBench evaluates agents that use terminal tools to work with audio and video files. The Terminus-MM harness adds multimedia perception to study how accessible evidence changes executable workflows and task outcomes.
Reorganizes webpage accessibility trees into functional regions through hierarchical decomposition and semantic abstraction. PageDigest supplies a persistent region-level observation to web agents, evaluated on WebArena.
Introduces Weblica, which combines HTTP-level capture and replay with website-grounded environment synthesis. The resulting interactive web environments support reinforcement learning and tests of navigation transfer and inference-time scaling.
Describes an Android agent architecture combining multimodal perception, working and personal memory, hybrid XML/visual grounding, and reusable skills captured from user interactions.
Evaluates agentic browsers against classical web attacks and visual social-engineering lures beyond indirect prompt injection.
Presents AutoFocus, a training-free grounding method that turns sampled coordinates and token uncertainty into spatial region proposals. Shape-aware crops and visual comparison select a target while retaining surrounding interface context.
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.
A11y-Compressor transforms linearized accessibility trees through modal detection, redundancy reduction and semantic structuring. Its OSWorld evaluation measures both observation-token reduction and task success.
This study compares specialized, computer-use and coding agents on scientific visualization workflows. It examines GUI, CLI, code and tool-call interaction modes, including how persistent memory changes performance across repeated trials.
Introduces WindowsWorld for professional workflows spanning multiple desktop applications, with generated tasks refined through human review. Intermediate subgoals support process-oriented evaluation of cross-application reasoning and execution efficiency.
AutoSurfer explores website actions breadth-first, retaining page and trajectory knowledge to guide task synthesis and trajectory refinement. It evaluates the resulting training data by fine-tuning a web agent on WebArena and examining synthesized-task diversity.
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.
AniMINT evaluates how VLMs perceive, explain and interpret UI animations using densely annotated videos. The study distinguishes basic motion recognition from higher-level animation meaning and probes motion, context and perceptual cues.
uxCUA learns to assess interface usability by choosing interaction flows, executing them and predicting a numerical usability score. Training uses interactive UIs with usability labels and human preferences, with evaluations on synthetic and real interfaces.
SnapGuard detects screenshot-based prompt injection through complementary visual-stability and recovered textual cues. Its evaluation compares detection quality and inference cost across attack and benign-page settings, rather than establishing universal runtime prevention.
Introduces DynamicGUIBench for applications whose interfaces change between actions, together with DynamicUI, an agent that processes screen recordings. The agent selects informative frames, filters inconsistent action reasoning and uses trajectory reflection for subsequent decisions.
Introduces Odysseys, a live-web benchmark of long, multi-site tasks derived from browsing sessions. Graded task rubrics and a rubric-score-per-step measure evaluate both partial progress and trajectory efficiency.
AutoGUI-v2 evaluates interface functionality, grounding and predicted interaction outcomes using hierarchical regions derived through VLM-human annotation. Its tasks span multiple operating systems and expose differences between functional localization, captioning and transition-logic understanding.
Provides an evaluation toolkit for joint safety, efficiency, and multimodal robustness analysis of operating-system GUI agents.
Studies a compact encoder-decoder grounding model and progressive training-data refinement for resource-constrained GUI interaction.
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.
Trains a GUI grounding proposer together with a visual critic that evaluates candidate locations rendered on screenshots. An adaptive reinforcement-learning schedule balances their development, using proposal diversity and critic feedback during training.
PlayEval measures whether generated GUI applications remain logically playable, using PlayTester to execute task-oriented GUI playthroughs. PlayCoder closes the loop between repository-aware generation, interactive testing and targeted code repair.
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.
AJ-Bench evaluates judge agents that actively gather evidence from search, data systems and graphical interfaces. Its annotated tasks and trajectories test information acquisition, state verification and process verification.
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.
Organizes security risks and defenses for persistent agent memory into six lifecycle phases and proposes five verifiable governance primitives.
Presents OpenMobile, which builds exploration-based environment memory to synthesize task instructions and switches between learner and expert policies during trajectory collection. The resulting data includes recovery behavior and is tested through mobile-agent training and benchmark-overlap analysis.
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.
DOMSteer translates help requests and live page context into reversible DOM changes such as control highlighting, contextual tooltips and layout reorganization. The work evaluates in-place GUI assistance with interface tests, use cases and a comparative user study.
AgentClick gives terminal-based agents a browser review interface for plans, drafts, execution traces and editable memory. A local server and skill plugin let users inspect proposed actions and provide structured feedback.
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.
Introduces UI-Zoomer, a training-free grounding method that uses candidate agreement and token confidence to decide when to crop and retry. Prediction variance determines a separate crop radius for each uncertain interface target.
UI-Copilot separates GUI execution from on-demand memory retrieval and numerical assistance. Its training procedure treats tool selection and multi-turn task execution as distinct optimization targets while keeping persistent observations outside transient context.
WebXSkill pairs parameterized action programs with step-level language guidance, mining reusable skills from synthetic browser-agent traces. A URL-based graph supports retrieval, and deployment alternates between direct program execution and guided native planning.
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.
WebAgentGuard separates prompt-injection detection from the operating web agent through a parallel multimodal guard. The guard is trained on synthetic visual/textual examples using reasoning-focused supervised fine-tuning followed by reinforcement learning.
Presents ClawGUI, a shared harness for GUI-agent reinforcement learning, benchmark evaluation and deployment on mobile devices. It combines parallel virtual and physical environments, process rewards, and hybrid CLI-GUI execution with persistent memory.
Presents WebForge, a pipeline that plans, generates, refines and validates self-contained interactive websites. WebForge-Bench uses these environments to evaluate browser agents across domains and controlled difficulty levels.
Introduces HealthAdminBench with GUI environments for an electronic health record, payer portals and fax handling. Expert-defined administrative tasks are decomposed into verifiable subtasks to compare local progress with end-to-end workflow completion.
EE-MCP studies application-dependent mixtures of GUI actions and MCP calls using generated environments, trajectory collection and filtered training. An experience bank extracts rules from trajectory comparisons for inference-time improvement, distinct from model distillation.
Introduces CORA, which calibrates a risk model to choose between executing and abstaining from mobile GUI actions, then diagnoses rejected actions to suggest interventions. Phone-Harm provides step-level safety labels for evaluating this safeguarding process.
Introduces ClawBench, 153 everyday tasks on production websites in 15 categories; an interception layer blocks each final submission so bookings, purchases and job applications are scored without side effects.
Presents MolmoWebMix, combining synthetic browser trajectories, human demonstrations and GUI perception data, and MolmoWeb screenshot-to-action policies trained from it. Browser benchmarks also test parallel rollouts with best-of-N selection.
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 Android Coach, which uses a learned critic to evaluate multiple candidate actions at each observed emulator state. Process rewards and a group-wise advantage estimator support online training, evaluated on AndroidLab and AndroidWorld.
Showing 361–420 of 1,202, newest first.
