已审阅并标注「效率」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。
Agent JIT compilation turns browser task descriptions into validated executable plans containing model calls, tools and parallel work. A cost-based scheduler and pre/postcondition protocol target latency and incorrect tool use in web-agent execution.
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.
TClone versions live GUI workspaces through snapshots, isolated forks, rollback and selective commit or merge. It separates fast branch creation from durable checkpoints using copy-on-write sharing, filesystem versioning and asynchronous checkpointing, with end-to-end agent-loop latency measurements.
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.
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.
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.
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.
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 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.
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.
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.
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.
PixelPrune removes pixel-redundant patches before the vision encoder, reducing both encoding and downstream language-model work. The training-free method evaluates lossless and controlled-lossy compression on document and GUI tasks.
MolmoPoint replaces textual coordinate generation with pointing tokens that select visual regions and refine them through subpatches. The general image/video method includes GUI-pointing evaluation on ScreenSpotPro alongside other visual grounding tasks.
Adapts a discrete vision-language diffusion model to single-turn GUI action and bounding-box prediction, using a hybrid masking schedule for coordinate geometry. Experiments across web, desktop and mobile datasets examine grounding accuracy and inference tradeoffs.
Studies visual token pruning for GUI navigation, separating reusable screenshot encoder embeddings from later language-model computation. The analysis examines foreground-background token allocation and spatial coverage when pruning historical screenshots.
MANA combines visual, temporal and exploration signals to guide navigation toward mobile advertisements. It evaluates ad detection and interaction cost on commercial smartphones, including obfuscated-ad cases.
Adaptive VLM Routing estimates GUI-action difficulty and small-model confidence to select a model under a predicted reliability threshold. Grounding and agent-routing benchmarks assess the cost–accuracy trade-off, with prior interaction context used for warm agents.
ActionEngine explores web applications to build an updateable state-machine memory of GUI states, operations and transitions. It generates complete task programs from that memory, using reactive repair to handle changed interfaces or missing knowledge.
CATTS allocates additional inference computation when an agent’s action votes show uncertainty. The paper compares voting and arbiter strategies and evaluates this selective scaling on WebArena-Lite, Online-Mind2Web and GoBrowse.
TreeCUA organizes GUI exploration trajectories as reusable trees and balances deeper tasks against broader coverage. Its data pipeline combines action verification, memory-backed recovery and a DPO variant that uses information from adjacent branches.
Agent Alpha uses step-level Monte Carlo tree search to reuse successful prefixes and recover from early GUI mistakes. Comparison-based evaluation and diversity-constrained expansion guide the search, with experiments on OSWorld.
VAGEN verifies GUI trajectories by first extracting readily available evidence and escalating to tool-based environment probes when needed. Evaluation on OSWorld-Verified and AndroidWorld studies verification accuracy and its cost.
GraphPilot builds app-specific knowledge graphs from explored pages, elements and transitions, then uses them to plan action sequences. It requests additional interface observations when uncertain and checks plans against graph transition rules before correction and execution.
CCPO couples GUI policy optimization with coordinate-guided compression of past visual observations. It aggregates targets across rollouts to focus retained context and uses distance-based advantages to train grounding and compression together.
FocusUI selects instruction-relevant screenshot patches while preserving their positional structure for GUI grounding. Its PosPad scheme replaces runs of dropped tokens with positional markers and is evaluated under reduced token budgets.
WebGym supplies realistic web tasks with rubric-based evaluation and an asynchronous rollout system for reinforcement learning. It studies how training on an agent’s own interaction traces scales with task coverage and transfers to unseen websites.
NestBrowse organizes browser actions in a nested framework that separates interaction control from page exploration. It evaluates this design for deep information seeking beyond snippet retrieval and direct URL fetching.
MAI-UI combines GUI navigation, native user interaction, and MCP tool calls in a family of policies trained through generated data and online reinforcement learning. The report also studies device-cloud routing and scaling of parallel environments and interaction budgets, with grounding and Android navigation evaluation.
Introduces iSHIFT, which combines latent reasoning with perception-control tokens to adapt GUI inference. The agent switches between detailed regional grounding and faster use of global visual cues.
HiconAgent trains GUI policies to use history through variable-length context sampling and paired policy updates with compressed and complete histories. Its compression retains past actions while removing historical observations, and an alignment loss couples the two branches.
Introduces AFRAgent, which uses adaptive feature renormalization to combine low-resolution visual embeddings with high-resolution details. The architecture is evaluated for smartphone GUI automation on Meta-GUI and AITW.
Introduces Prune4Web, where generated scoring programs filter DOM elements according to decomposed subtasks before grounding. An annotation pipeline and two-turn training procedure coordinate the planner, programmatic filter and grounder.
Introduces FaraGen synthetic web trajectories, a compact screenshot-based action model, and WebTailBench for underrepresented browser tasks.
OSGym supplies full operating-system sandboxes with GUIs for parallel computer-use data collection and training. It combines per-replica state management, hardware-aware orchestration, copy-on-write virtual-machine disks, and layered failure recovery.
Introduces GUI-AIMA, which supervises instruction-relevant attention over visual patches instead of directly training coordinate text generation. Patch-level grounding can be combined with an optional zoom stage and is evaluated across GUI localization benchmarks.
OpenPhone trains a compact mobile policy with supervised learning followed by GRPO on synthetic GUI data, incorporating reasoning and memory management. Its device-cloud design handles routine work locally and delegates complex subtasks, with online evaluation on AndroidLab.
WebServ combines copy-on-write container environments with a compact DOM-derived browser interface for training web agents. Network-aware waiting supports action execution in single-page applications, with evaluations and reinforcement-learning experiments on WebArena-Lite.
Introduces CORE, which divides mobile UI hierarchies into semantic blocks and shares planning between local and cloud language models. Local ranking limits the screen information sent for cloud-side element selection, with a multi-round mechanism for correcting inadequate context.
Encodes GUI trajectories into compact continuous embeddings that can be retrieved directly into an agent’s input. An automated environment-discovery and rollout pipeline expands this memory for long tasks and unfamiliar interfaces.
WALT explores website functionality and converts demonstrated interactions into validated, reusable tools combining DOM actions, URL navigation, and agentic fallbacks. It evaluates this expanded browser action space on WebArena and VisualWebArena.
GUI-KV compresses vision-language agents’ attention caches using visual-token saliency and redundancy between screenshot frames. The method requires no retraining and evaluates how cache budgets affect GUI action accuracy and computational cost.
Ferret-UI Lite develops a 3B GUI agent using mixed real and synthetic data, reasoning with visual tools at inference, and reinforcement learning. It evaluates grounding and navigation across mobile, web, and desktop settings, including AndroidWorld and OSWorld.
DART separates GUI reinforcement learning into asynchronous environment, rollout, data-management, and training services. It curates experience by task difficulty, successful trajectories, and step uncertainty while correcting rollout-policy mismatch, with evaluation on OSWorld.
MobileRL adapts GRPO to mobile GUI training with difficulty-dependent replay of successes and filtering of failure curricula. A shortest-path reward adjustment accounts for task length, with evaluations on AndroidWorld and AndroidLab.
Succeed or Learn Slowly applies direct off-policy updates to successful interactions and more conservative regularized updates to negative examples. Successful Transition Replay prioritizes successful experience during policy learning, with evaluation on AndroidWorld.
Presents MobiAgent, combining MobiMind agent models, the AgentRR acceleration framework, and MobiFlow evaluation. It also introduces an AI-assisted data-collection pipeline for mobile GUI tasks.
Introduces CogniWeb, which switches between fast reactive processing and slower deliberative reasoning according to task complexity. WebArena experiments examine this combination of learned behaviors and online planning.
D2Snap reduces serialized DOM observations while preserving actionable elements and features that distinguish their uses. The study compares reduced DOM inputs with screenshot-based observations for a web agent on tasks sampled from Online-Mind2Web.
CoAct-1 delegates subtasks between a GUI operator and a programmer that executes Python or Bash. Its OSWorld evaluation studies a hybrid action space in which scripts handle suitable file and data operations while visual interaction handles other steps.
MMBench-GUI evaluates interface understanding, element grounding, task automation, and collaboration across six named desktop, mobile, and web platforms. It introduces the Efficiency-Quality Area metric for examining execution efficiency alongside task quality in online GUI automation.
MobileGUI-RL constructs an online training curriculum through agent exploration and task filtering. It adapts GRPO with trajectory-aware advantages and rewards for task completion and execution efficiency, evaluated on three online mobile-agent benchmarks.
Studies compute allocation for web-agent post-training with teacher-generated supervised fine-tuning followed by on-policy reinforcement learning. It samples 1,370 training configurations and uses bootstrap analysis to examine hyperparameter sensitivity on WorkArena and MiniWoB++.
SimpAgent prunes distracting interface elements with visual masks and compresses interaction history using consistency-based guidance. The study evaluates these two forms of context reduction during GUI navigation in web and mobile environments.
LineRetriever uses a language model to select observation lines that matter for upcoming web-navigation actions. It reduces DOM or accessibility-tree context by considering the planning horizon instead of relying solely on semantic similarity.
ZonUI-3B trains GUI grounding with a curated 24,000-example collection of mobile, desktop, and web screenshots. Its two-stage procedure first learns across platforms and then specializes on high-resolution images, with ablations on data diversity and redundant examples.
OSWorld-Human adds human-determined trajectories to OSWorld to evaluate the efficiency of desktop computer-use agents. The study compares 16 agents and analyzes model-call latency, changes in step duration during task execution, and excess action steps.
第 61–120 条，共 158 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。
