从 GUI 定位到智能体训练、评测与安全，每篇都可追溯到原始论文。
阅读英文研究综述 ↗ · 方法、评测与人的控制权；英文工作稿附有逐篇阅读记录。
条目经过 AI 辅助的一手来源编辑审阅，实验未独立复现。发表日期保留来源记录的精度。下方资源链接来自已审阅的介绍或文档关系。 收录方法 ↗
OpenForgeRL connects stateful agent harnesses to reinforcement-learning infrastructure through a model-call recording proxy and isolated rollout containers. Its validation includes browser and computer-use agents, alongside tool-oriented agents, and examines how harness choice affects learned reliability.
Measures how autonomous LLM browser agents interact with commercial bot-management and anti-automation defenses on the web.
Introduces SEE, which first explores mobile interfaces to build a screen-and-element transition graph, then composes task trajectories through graph planning and controlled sampling. It studies coverage, loop avoidance, and transfer after training agents on the generated sequences.
Introduces Sidekick, a communication interface that uses ambient status cues, multimodal action summaries, and visual and verbal explanations of reasoning to help people monitor computer-use agents. A user study examines progress awareness, error traceability, and multitasking across background and foreground interaction.
Introduces EvoGUI, which derives temporal ordering, inverse action/value prediction, and successor-discrimination probes from recorded GUI trajectories. Its web-derived benchmark isolates transition understanding from the other abilities involved in end-to-end task success.
Introduces ParaGUIBench, with tasks and efficiency measurements for coordinated GUI agents operating separate desktop instances over a shared file system. Its ParaGUI planner-worker system decomposes tasks and dispatches concurrent subtasks for comparison with serial execution.
Uses instruction screening and a safety-augmented world model to assess likely consequences of mobile GUI actions before execution.
Plover exposes GUI-agent plans and replanning as persistent artifacts that users can inspect and edit during execution. Its planner–executor design supports localized correction through plan edits, language guidance and screenshot-grounded interventions, studied through failure repair and workflow scenarios.
Combines GUI perception data, instructional videos, real and simulated devices, and planning with reflection to develop a mobile agent.
Tactile turns accessibility, OCR and visual evidence into explicit desktop targets with executable actions and verification cues. Its execution loop chooses semantic actions or coordinate fallbacks while retaining provenance for replay and diagnosis.
Studies a controlled setting where GRPO fails to improve a strong small web agent and can reduce text-track performance. Matched reachable-reward tasks and component grafting distinguish limited optimization headroom from learning-rate-dependent failure mechanisms.
KnowAct-GUIClaw combines a host planner with pluggable GUI subagents, attributable experience memory and an evolving skill library. It uses user profiles and feedback for task decomposition and evaluates transfer across Android, iOS, HarmonyOS and Windows.
StructAgent maintains compact task state and regulates long-horizon execution through verified transitions and progress checkpoints. Its workflow supports evidence-based completion and targeted recovery, with desktop computer-use evaluation and a separate Minecraft extension.
Scales computer-agent reinforcement learning through verifiable task generation, capability-aware rollout allocation, and segmented visual context.
The Entry–Propagation–Recovery framework traces where conflicting memory first changes an agent action, how the effect propagates and whether recovery occurs. WebArena and the controlled MemTrapBench separate memory-consumption failures across these stages.
MAG jointly evaluates browser task actions and user-facing guides grounded in screenshots through element marks or pixel coordinates. Its harness covers annotation, live evaluation and joint metrics, with an expert-trajectory-augmented GRPO training study.
EvoCUA-1.5 trains multi-turn desktop agents through online reinforcement learning in executable sandboxes. It combines step-level policy optimization, task filtering and calibration, an adaptive curriculum, and asynchronous rollout handling with staleness control.
WebRetriever evaluates navigation, knowledge-assisted interaction and end-to-end information-extraction workflows under separate protocols. Its NavEval judge uses richer interaction context to diagnose cases where navigation success does not translate into useful task completion.
Untrusted Content Masking uses DOM structure to redact untrusted webpage regions before agent observation. A sandboxed interaction interface separates privileges so browser actions can proceed under the proposed trust-boundary model without exposing protected content to the planner.
This study examines agent data injection, where untrusted content impersonates trusted metadata or tool-context data. It reports arbitrary-click vulnerabilities in web agents alongside separate coding-agent attacks, highlighting trust-boundary failures beyond instruction injection.
Introduces Uni-GUI, a collection of executable desktop and mobile trajectories, and UI-MOPD, which trains a shared policy from its own rollouts using platform-specific teachers. The distillation procedure routes each visited state to the appropriate teacher to preserve distinct interaction conventions.
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.
Introduces NeXUI, an assistive UI-agent benchmark that evaluates navigation together with interface-grounded explanations and user oversight in recorded interface states.
Analyzes cross-app privilege escalation and visual attack surfaces introduced by third-party VLM-powered mobile agents.
DigitalCoach records expert–novice software coaching conversations aligned with screen and input events. Automated and interactive evaluations examine how model coaches differ from humans in explanation, error diagnosis and visual grounding when helping a learner operate software.
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.
LUMOS converts native accessibility metadata and browser UI structure into semantic interface blueprints with stable element identifiers and available actions. It combines live pointer-to-element grounding with an observe-act loop constrained to visible-UI primitives.
Evaluates web agents against anti-bot mechanisms on instrumented honeysites and analyzes network, HTTP and browser fingerprints. The study compares resistance to blocking with the ability to distinguish agents, humans and agent implementations.
InnerZoom preserves target-region cues from intermediate decoder layers and reinjects a compact evidence state during coordinate decoding. Its single-forward GUI grounder is compared with crop-and-rerun zooming and matched supervised/RL baselines for accuracy and inference cost.
GUICrafter first learns grounding from large collections of unannotated screenshots and webpages, then uses a smaller high-quality dataset for reinforcement-learning calibration. The curriculum targets reduced dependence on manually annotated GUI training data.
Introduces extended computer-use workflows with realistic artifacts, changing information, hidden state, and partial-progress and safety assessments.
The Agent-Computer Observation Interface separates continuous perception from discrete computer actions using captured frames, audio transcription and persistent visual narration. DynaCU-Bench tests these components on dynamic browser tasks and static controls, with model-specific ablations showing that additional observations can also hurt.
TUA-Bench evaluates terminal-use agents on general digital tasks, including document editing, email, web information seeking and specialist software workflows. Tasks run in real terminals with deterministic setup and execution-based scoring.
PEEU explores web environments and turns hindsight experience into aligned high-level planning supervision. A hierarchical task-decomposition analysis examines why atomic-skill competence alone does not ensure planning generalization across websites.
WinDOM obtains bounding-box supervision from a Windows 11 web reimplementation driven through Playwright, and studies rejection-sampling distillation for small GUI grounders. It treats the depth of the supervised cold start as a GRPO choice, comparing same-size EMA and larger same-family teachers.
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.
This short paper develops an explorer agent that expands from a demonstrated GUI task to discover requests likely to reach user-sensitive screens. The proposed exploration supports identifying situations where task execution should be handed back to the user.
VisCritic compares before/after screenshots through change-aware visual features and an action-conditioned critic. Weakly supervised samples from existing traces train judgments of action success, progress and error type for GUI-agent verification.
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.
AgentCIBench evaluates whether computer-use agents disclose information in ways that violate its original context. Deterministically scored scenarios test nearby sensitive UI content, ambiguous requests and recipient mismatches, including end-to-end agent execution.
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.
Introduces macOS tasks spanning GUI and command-line interaction, with deterministic checks and intermediate-progress scoring for application workflows.
ChainWorld composes compatible atomic OSWorld tasks into multi-objective desktop chains while preserving their evaluators. It compares giving the full sequence at once with revealing goals turn by turn, exposing different state and session-management failures.
Builds a logging framework that combines network/protocol fingerprints with browser interaction behavior to attribute web-agent traffic. A decision-tree classifier is evaluated across six agent implementations, human browsing and conventional crawlers.
Builds browser-agent training trajectories with live and synthetic environments, solvers, user simulation, and complementary correctness and efficiency verifiers.
Segments and clusters GUI trajectories into inspectable skills, then trains a skill-aware policy from those annotations. Its diagnostic results distinguish readable source-domain clusters from reliable transfer and report limited or absent downstream gains.
MobileForge couples real-app exploration and generated tasks with hierarchical feedback for mobile-agent adaptation. Its optimizer turns trajectory outcomes, step feedback and corrective hints into contextualized GRPO updates, with evaluation on AndroidWorld and a GUI-only MobileWorld split.
MemGUI-Agent makes context management an action chosen by the same policy that controls the mobile UI. Structured history, UI-state and recent-step fields are trained from annotated trajectories to retain cross-app facts while limiting context growth.
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.
HANSEL extracts evidence pages and snippets from browser-agent traces and preserves relevant page state in interactive verification views. It flags untraceable answers and evaluates evidence selection and human verification effort separately.
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.
This study compares general coding agents using mobile command-line access with reproducible GUI agents on AndroidWorld and MobileWorld. It also supplies oracle CLI analyses and a separate task suite for operations such as bulk changes, filtering and hidden device state.
VISUALSKILL organizes application-specific text and figures into hierarchical skills retrieved on demand through an MCP tool. Skills combine authored documentation with live UI exploration, and matched multimodal/text-only comparisons test their utility for interface identification and workflow-state verification.
Proposes quality-aware self-distillation for GUI coordinate generation under student-produced prefixes. A correctness-aware gate suppresses coordinate-token supervision that cannot reach the target box, while teacher-probability scaling adjusts the remaining signal strength.
PreAct compiles successful screen-interaction runs into guarded state-machine programs for repeated tasks. It validates programs from clean states before storage and returns control to the agent when replayed screen expectations fail, with tests across mobile, desktop and web.
ProCUA-SFT generates tasks and synthetic trajectories on live desktops seeded with document content, checking task preconditions before rollout. The resulting step-prefix examples are formatted to match inference context and used for supervised GUI-agent fine-tuning.
LabOSBench evaluates GUI agents on browser-based scientific instrument simulators with configurable tasks and execution-based checks. Its subtask and end-to-end tests examine feedback-driven adjustment and long workflows without controlling physical instruments.
MyPCBench tests personal-assistant tasks on a Linux desktop with simulated web applications and data for a single fictional persona. A shared computer-and-shell interface enables evaluation of cross-application workflows that depend on personal context and account-like state.
第 241–300 篇，共 1,200 篇，按发表时间由新到旧排列。
