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。
Tandem serves hybrid GUI-agent models on Android NPUs using divergence-point state checkpoints, rollback-aware speculative decoding and idle power release between request bursts.
VeriCUA trains a Judge to select verification checks and stopping decisions while a frozen Probe inspects the post-task GUI. Single-attempt reinforcement learning balances verdict correctness and Probe use.
Uses weak-agent uncertainty to decide when to escalate to a stronger model and monitor its subsequent task execution.
ECHOES learns recursive textual task memory through privileged screenshot-history distillation and GRPO, then predicts GUI actions from the current screenshot and memory.
Introduces STP-GUI, which selects spatially coherent screenshot windows conditioned on the instruction to reduce visual tokens while retaining useful interface structure.
Adds decision heads to a frozen vision-language generator for selecting web actions and targets, with a confidence gate that falls back to text generation.
Introduces WebKD, which repairs failed student web-agent trajectories with minimal teacher edits and trains on corrections that preserve the student's reasoning style.
Introduces CarryOn, which maintains evidence-linked task objects and remaining-work context across history compression, with evaluations on desktop tasks and web research.
Replaces future-screenshot generation with a shared latent representation for GUI action selection and action-conditioned transition prediction. An action-aware contrastive objective encourages the transition representation to retain discriminative action-induced changes.
Introduces Web$hell, an interface that compiles live pages into persistent information and action objects so agents can gather evidence, navigate websites, and reuse acquired content through shared operations.
Evaluates Jev as a bounded action selector inside three AndroidWorld agents and two text-based ALFWorld agents, separating decision savings from task success, recovery and resources per delivered outcome.
Introduces TeleTune, which edits a textual skill library using held-out action prediction on offline software-use logs with missing goals or interleaved tasks. The learned workflows also guide retrieval of demonstrations for subsequent web-agent execution.
Introduces ReLaViS, which performs recurrent visual search inside a grounding model by repeatedly querying screenshot tokens and feeding the resulting spatial evidence back into the model. Coarse-to-fine supervision is constructed from existing interface-element annotations.
Studies a minimal terminal-based execution framework that lets digital agents construct tools and manage context across browser, desktop, and CAD tasks.
Introduces PACE, a protocol and online decision rule for deciding when repeated GUI procedures are worth compiling into programs. It accounts for failed compilation attempts, measured per-use savings, uncertain reuse, and execution costs under an explicit cumulative budget.
Studies how to route different parts of GUI experience into model weights or runtime context rather than treating whole trajectories as a single unit. It separates locators, procedures, state facts, and lessons, then predicts their destination from recurrence and dependence on the current state.
Standardizes virtual machines, interaction pipelines, and agent interfaces to compare computer-use speed, cost, and task performance.
Learns reusable computer-use policies that encode stable workflow logic in programs while delegating grounding and state checks to neural models. Policy iteration executes variants, diagnoses failures, revises the program, and adds verification before state-changing actions.
Introduces PhoneCLI, which explores mobile app interfaces and compiles navigation paths into reusable commands without internal app APIs or instrumentation. At execution time it verifies compiled paths and falls back to a visual agent for open-ended interaction or failed navigation.
Introduces CUA-Sandbox, which separates each rollout’s mutable state from shared initialized software runtimes. State-scoped execution and transactional resets or branches aim to reduce training-environment costs while retaining original interfaces and evaluators.
Introduces PastForward, which reuses earlier output sequences as verified decoding proposals and starts future action computation from predicted GUI transitions. Screen checks determine whether speculative computation and cached model states can be retained.
Introduces REBASE, a device–cloud protocol for repairing cached GUI experience after app updates. It combines cloud replay, device-side step verification, and progressively richer failure reports to produce repairs tied to application versions.
GUI-Hopper combines verified Android deeplinks with GUI actions, using landing-screen descriptions, hybrid trajectory training and reinforcement learning to select navigation shortcuts and recover through GUI control.
Separates occasional visual-language planning from frequent lightweight action selection over an accessibility-tree-based action space.
Trains a mobile planner with supervised trajectories and competence-adaptive RL, then couples typed tool execution with memory, skills and offline model-harness refinement.
PACE weights coordinate-digit entropy by place value to estimate confidence in GUI clicks from one forward pass. It evaluates correctness ranking and selective accuracy against repeated-sampling alternatives on ScreenSpot-Pro and ScreenSpot-v2.
RankGround trains a lightweight reranker to select a promising screenshot crop before one VLM localization call. Ranking supervision uses containment and boundary-aware augmentation, with pointwise and listwise training stages.
EchoPath turns validated GUI trajectories into parameterized replay procedures with explicit runtime preconditions. Image-based target matching adjusts stored coordinates, while ambiguous steps fall back to bounded repair or fresh planning.
AnchorGUI compares expected and observed UI transitions and selectively retains screenshots when they disagree. This asymmetric memory supports immediate correction and focuses cross-attempt experience distillation on likely failure steps.
AutoTailor compiles web trajectories into reusable browser-automation APIs, then filters and dynamically revises the capability inventory. WebArena experiments distinguish API-only operation from an agent that can fall back to ordinary reasoning and browser actions.
OdoBot builds a model of web-application behavior from successful task demonstrations and uses it to guide task execution. Its Canvas LMS evaluation compares token use and completion with existing web agents.
LLaDA-UI adapts block-wise diffusion generation to a vision-language GUI agent through multimodal pretraining and GUI supervised fine-tuning. It evaluates grounding and navigation across multiple platform settings.
TRACE orders screenshot evidence for monotonic visual-token reduction using layout priors, instruction relevance and feature novelty. Spatial token reserves and incremental KV contraction aim to preserve useful evidence for future GUI targets without repeatedly encoding retired frames.
GraphDroid combines clustered memory of visited GUI states with asynchronous test-intent generation. A hybrid executor reserves LLM calls for complex intents and uses heuristics for simpler ones, evaluated through Android app coverage, bug discovery and crash detection.
DroidTool generates application-state tools through proposal, implementation, testing and repair. Relational tests connect the tools’ preconditions and effects, and the resulting tools are evaluated as additions to Android GUI agents.
Proposes an online teaching framework that decides which web-agent episodes warrant a stronger teacher and which trajectory turns merit training. A solvability gate and score-guided trimming aim to preserve first-attempt performance while reducing teaching and training costs.
Introduces CausalCache, which selects archived screenshots to restore alongside text summaries under a fixed visual-memory budget. A history-gated adapter processes restored images, with evaluation distinguishing standard step limits from longer-horizon and mobile-transfer settings.
Introduces LACL-GUI, a contrastive learning objective that distinguishes trajectory quality within both successful and failed GUI executions. It favors concise successes and uses divergence from successful trajectories to differentiate failures during policy optimization.
ComponentBench evaluates short interactions with canonical web UI components using verified tasks and matched human reference trajectories. Its diagnostic pipeline compares observation/action representations and analyzes structural difficulty, failure patterns and interaction efficiency.
Trains HTML-element relevance filters on Mind2Web trajectories and transfers them to WebArena to reduce irrelevant browser context. It also evaluates a zero-shot ColBERT retriever and documents baseline web-agent failure modes.
Measures the complementarity, cost and rerun variability of six browser observation modes on WebArena and VisualWebArena. Its routing experiments distinguish an oracle cost opportunity from policies that fail to robustly outperform a fixed observation mode.
StepReflect predicts structured per-action reflections from transition specifications and paired visual evidence. The paper distinguishes offline transition-label accuracy from online task success when the reflection model is integrated into mobile agents.
This study separates tool-adoption behavior from task competence in hybrid GUI–MCP agents. It tests tool-use incentives and screenshot-history compression, showing why changed action preferences alone need not improve held-out accuracy.
Examines inference-time scaling for local computer-use models by varying context, step budgets, decomposition, and parallel execution. OSWorld experiments track both task performance and changing failure modes, exposing cases where extra computation extends errors or adds coordination overhead.
Introduces Adaptive Anticipatory Policy Trees, which precompute guarded actions and deadlines during idle periods so transient GUI events can be handled without new text generation. Paired experiments examine when this prepared execution helps and when reactive control remains preferable.
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.
Scales computer-agent reinforcement learning through verifiable task generation, capability-aware rollout allocation, and segmented visual context.
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.
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.
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.
WebChallenger builds a hierarchical DOM-derived page representation for selective observation, reusable site maps and compound actions. The framework combines these mechanisms without model fine-tuning and evaluates generalization across multiple web-agent benchmarks.
SUPERBROWSER combines screenshot-based target proposals with separate orchestration, planning and action roles. A bounded ledger and context-eviction process retain recent actions, task facts and checkpoints during browser execution.
Proposes Signal-Driven Observation: a separate call selects relevant DOM elements and selectors, and a detector requests new observations only after meaningful browser events or failures. The paper frames observation timing and compression as an architectural research problem for long-horizon web agents.
AsyncWebRL overlaps web-agent rollouts, updates and policy refresh while managing screenshot handling and a persistent rollout pool. It also replaces trajectory-length normalization in multi-step GRPO to reduce verbose failed trajectories, with throughput and out-of-distribution evaluation.
LivingScreen evaluates agents in a browser-based short-video environment where observing content consumes time and affects subsequent actions. Its task suite scores accuracy together with information efficiency to diagnose over- and under-observation on continuously changing screens.
STaR-KV compresses GUI-agent KV caches by calibrating visual-token importance across attention subspaces and interaction time. Its training-free scoring combines spatial specialization, temporal redundancy and an adaptive temperature, evaluated at matched cache budgets.
MACU represents computer-use objectives as a dependency graph whose manager dispatches parallel subagents and revises tasks as new information arrives. Desktop and web experiments examine coordination, information handoff and task-completion time.
UI-KOBE explores mobile applications to build graphs of UI states and executable transitions. A lightweight agent then matches its current screenshot to the graph and chooses a transition, local action, completion signal or fallback action.
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.
Analyzes inefficiency in VisualWebArena trajectories and distills reusable online skills to reduce repeated actions and token overhead in multimodal web agents.
第 1–60 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。
