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。
ScopeMem conditions latent-memory prefixes on a frozen GUI agent's initial action and trains them to correct errors while preserving correct decisions. Its evaluation separates action repairs from harmful revisions.
WebMix uses a scout-produced interface manual to route mixtures of source-trained LoRA adapters for web agents, adapting to target interfaces without target-time gradient updates.
Introduces GraceBench, using application transition graphs and recorded interaction evidence to diagnose deviations, loops, and recovery that final task-success scores can hide.
Introduces CLASHarness, generating verifiable computer-use tasks with conflicting agent reports, shared-state constraints, and coordination requirements to evaluate conflict detection and resolution.
Introduces MHVF, a training-free grounding method that explores several candidate screen regions, verifies them in local views, and uses those checks to guide further search.
Analyzes recurring distractor clicks in GUI grounding and introduces ConClick, a contrastive training method built from target–distractor pairs. Its ReClick inference procedure revisits candidate regions and compares competing predictions.
Introduces MisVisAgentBench, a paired executable benchmark that changes chart evidence while holding web tasks fixed, to trace how misleading visualizations affect agent decisions and downstream actions.
Introduces PAIR for controlled evaluation of mobile GUI agents across user-conditioned application states and RePAIR for training from cross-user subgoal outcomes. The study examines failures when personalized interfaces change which items an agent must identify.
Compares ways to condition a small GUI grounding model on action type under matched training conditions on Android in the Wild. The study traces apparent gains to a coordinate-serialization choice and reports limited evidence that the gains persist with predicted action types.
Describes WebFovea, a vision-based live-web agent whose execution loop checks action parsing, browser effects, feedback accuracy, and observation delivery. The report analyzes failures such as coordinate mismatches and silent control interactions and documents corresponding harness changes.
Improves GUI-agent execution code by analyzing action effects and repeated-task failures, then validating bounded changes with frozen model weights.
Combines structured GUI exploration with training designed to tolerate imperfect retrieved experience. It builds state-transition graphs and synthetic task trajectories, then injects realistic forms of interface drift so agents learn when to rely on or reject prior guidance.
Introduces Talk2Agent, a benchmark of spoken instructions for computer-use tasks that measures whether voice interfaces preserve task-critical information. It pairs human speech with an execution-free intention-based metric and checks how that metric relates to downstream task completion.
Studies repetitive, unproductive actions in desktop agents through their internal activation states. It proposes Reset, Reroute, Restore, a context intervention that temporarily resets an agent’s trajectory and later restores it to help escape repeated ineffective behavior.
Presents PrecogUI, which retrieves prior anomaly and success patterns and predicts symbolic UI transitions before choosing actions. It also introduces InterfereBench, generated for evaluating long GUI workflows with disruptive events and recovery requirements.
Describes a WebRetriever agent that combines semantic webpage operations with visual localization when structured representations are insufficient. Hierarchical context management and fault-aware execution support processing tasks across concurrent browsers.
Verifies web tasks through typed interface events and generates controlled UI variations while preserving instructions and success conditions.
Introduces Veer, a runtime defense that changes task-relevant web state when an otherwise valid proposed action would produce an unauthorized consequence. It plans a prospective intervention and grounds and verifies its execution before the base agent continues.
Defines a test-time adaptation setting with one counted attempt per arriving GUI task and no ground-truth feedback. Its SOLO method filters or relabels completed episodes, then updates a small adapter from a short window of the agent’s own experience.
Introduces AnyAppBench, a live Android benchmark that holds user goals fixed while changing the application used to achieve them. It combines category-controlled task pairs with failure annotations to distinguish task understanding from application-specific execution skill.
Introduces GUITAR, which groups visually different screens into functional states and analyzes agent failures over a state-transition graph. It uses the resulting bottlenecks to diagnose recurring mobile and web errors and evaluate targeted guidance.
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.
Moves controlled target icons across high-resolution interfaces to diagnose spatial blind zones in grounding models and study targeted augmentation.
Introduces CAVEAT, a controlled marketplace benchmark for testing whether delegated agents preserve user objectives when interfaces favor other outcomes. It diagnoses decision failures and studies harness interventions and targeted training to improve resistance to marketplace steering.
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.
BreakingWeb pairs solvable browser tasks with controlled environment interventions that preserve the goal, target and backend success criterion. The paired evaluations isolate robustness failures, including agents declaring completion without the required state change.
Studies how automatic defaults and reflective social-influence cues affect GUI-agent shopping decisions under different reasoning configurations. A randomized simulated-shopping experiment finds that reasoning changes the type of susceptibility rather than uniformly removing interface influence.
ERPBench evaluates screenshot-only agents in enterprise software using database values to check the business records produced by their actions. It pairs these state-based checks with a harness for human approval and analyzes cases where visible completion masks incorrect persisted data.
EvoSkill-GUI revises reusable skill files from execution feedback through an executor–critic loop. It studies how training-free updates to localization, recovery and procedure information help subsequent mobile and desktop tasks.
Investigates mismatches between what humans see on a mobile display and what agents receive through screenshots or accessibility metadata. An automated evaluation framework embeds instruction-agnostic perturbations into functional APK clones and compares static and dynamic agent deviations with human noticeability.
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.
Confidence-Anchored Learning uses coordinate-token confidence to filter pseudo-supervision for GUI grounding. Its negative-only variant performs label-free test-time updates to reduce reliance on potentially incorrect positive examples.
AgentHijack evaluates whether visual patch effects propagate from screenshots through model outputs and action parsing into real environment changes. Its controlled web-page experiments distinguish intermediate attack signals from verified end-to-end consequences.
Introduces activation-conditioned gradient control for GUI agents learning successive applications. It protects inactive application-specific neurons, constrains updates to shared neurons, and accumulates a compact historical state to balance adaptation with retention.
ElderBench evaluates mobile GUI agents on smartphone requests elicited from older adults, including indirect and underspecified language. Online and offline tests, instruction normalization and linguistic analyses examine how these request patterns contribute to failures.
CONFLICTGUI evaluates termination under instruction conflicts and mismatches between instructions and interface context. CONFLICTGUARD couples feasibility checks with conditional action modulation to reduce unnecessary execution while tracking normal-task behavior.
Predicts web-agent failure risk from observable trajectory prefixes using cross-step behavior and black-box consistency queries. Key-step labels preserve valid early prefixes of failed runs, and evaluation tests early intervention and transfer across website categories.
Introduces Test-Time Environment Decomposition, which uses trial interaction to break complex web observations into simpler modules. A label-free adaptation procedure learns from these sub-environments and studies whether their gains compose in full web tasks.
Introduces AnTrap, which injects task-solvable runtime anomalies into Android GUI-agent trajectories using a layered failure taxonomy. It compares model robustness and adversarial reinforcement learning to distinguish recoverable interface traps from persistent contextual failures.
Introduces Evidence-First Reflection, which separates extraction of action-induced screen changes from judgment of whether an operation achieved its objective. It filters visual differences into relevant evidence before a desktop agent’s reflector decides how execution should continue.
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.
SynWeaver explores a website to build a map of page states and executable interactions, then learns website priors for task proposals. Joint task/trajectory synthesis revises inconsistent pairs and verifies or repairs the resulting supervision.
MobileJudgeBench compares completion judges against human labels on mobile-agent trajectories. It studies screenshot-sampling baselines, model-dependent error profiles and how judge quality relates to agent rankings and reinforcement-learning rewards.
Evaluates interactive GUI agents under varied realistic interface noise, comparing task completion, safety, and behavior with clean-environment runs.
Extends AndroidWorld with controlled state, transition, and action perturbations to test mobile-agent robustness, and studies a training-free introspective recovery mechanism.
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.
This permission-literacy study tests whether mobile GUI agents grant only task-necessary permissions. Controlled changes to task context and requester identity expose authorization biases and the trade-off between suppressing unnecessary and legitimate grants.
Introduces LoginTrap, a black-box indirect-injection evaluation in which attacker-controlled page context and login flows induce unauthorized authentication. It measures end-to-end private-data leakage across web-agent backbones, architectures and defenses.
MissClick designs white-box attacks around the numerical effect of coordinate digits on executable GUI clicks. Separate untargeted displacement and targeted place-weighted objectives are evaluated across desktop, web and mobile grounding settings.
CUADebug combines a failure taxonomy and human-annotated OSWorld failure trajectories with an agent that locates causal steps and proposes corrections from screenshots and action traces. It evaluates root-cause diagnosis separately from controlled task re-execution and failure recovery.
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.
Audits failure-scored computer-use trajectories to distinguish agent errors from broken tasks and evaluator false negatives. It organizes reliability issues across task construction, trajectory evidence, scoring, and reporting, then derives corresponding evaluation design practices.
Introduces FCPAgent, which attaches confirming and falsifying evidence to each skill-grounded plan step. A plan–test–repair loop checks actions and observations, then localizes contradictions so execution, skills, or plans can be revised at the appropriate level.
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.
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.
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.
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.
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.
第 1–60 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。
