Source-reviewed research and assets tagged with Reward models.
Membership follows reviewed tags; it does not establish a dependency between neighboring records. Dates preserve source precision. Annotations retain their original language.
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.
Synthesizes resettable workspaces with files, application observations and skills; repairs environment incompatibilities and trains long-task agents using verifier-linked turn credit.
Introduces SURFPRM, which turns web demonstrations into an interaction-element graph for synthesizing grounded negative actions. Its contrastive preference data trains process reward models to distinguish plausible on-page alternatives and guide downstream trajectory search.
Trains discriminative web world models to distinguish the true next state from outcomes of alternative actions. Branching web trajectories support training, followed by tests of predicted-state matching, process-reward action ranking and end-to-end WebArena-Lite selection.
Introduces AdaptRubric, which retrieves reusable criteria for a GUI task family and then generates instruction-specific checks for concrete values and constraints. It evaluates these adaptive rubrics in offline reward assessment and online agent optimization.
Introduces GSAR, which synthesizes GUI tasks and successful goal states, then annotates task-relevant interface elements as reward anchors. These anchors support offline trajectory verification and online policy training.
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.
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.
Introduces OSReward and its harder and multidimensional variants for assessing judgments of computer-use trajectories. It also releases a reasoning-annotated judgment corpus and trains reward models, examining reliability, leniency bias, and evaluation cost.
Introduces an Interactive Reward Agent that proposes completion conditions and verifies them using system, application, and GUI tools after execution. GUI-RewardBench provides annotated desktop trajectories for evaluating these evidence-based judgments and their use in reinforcement learning.
Introduces SeekJudge, which iteratively searches and analyzes GUI trajectories to assess instruction fulfillment. The work combines dense judgment-data synthesis, a step-label benchmark, and reward-server scheduling to study accurate and practical reward feedback for agent training.
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.
HiViG combines a compact history of completed actions with a visually grounded critic that checks proposed coordinates before execution. Its evaluations separate history-related planning failures from errors caused by targeting the wrong screen element.
StainFlow tracks visually verifiable task entities through GUI trajectories to identify phases and retrieve focused evidence windows. Its process-reward model is tested for both online policy learning and trajectory-completion judgments.
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.
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.
Trains IntentScore on offline desktop GUI interactions to assess candidate actions using both their state relevance and planning intent. The reward model is evaluated as an action reranker for Agent S3 on OSWorld.
Describes a web-trajectory verifier built around structured rubrics, separate process and outcome judgments, and explicit treatment of uncontrollable failures. CUAVerifierBench supplies human labels for evaluating these choices and screenshot-based context management.
The framework separates inference-time subgoal planning from MiRA, a reinforcement-learning method that supplies milestone rewards. WebArena-Lite experiments evaluate these mechanisms for long-horizon web navigation.
Presents OS-Themis, a multi-agent critic that divides GUI trajectories into verifiable milestones and reviews their evidence before assigning an outcome. OGRBench evaluates GUI rewards across platforms, with AndroidWorld experiments examining training and trajectory filtering.
Builds ExeVR-53k from execution-video, instruction and reward triplets, including negative instructions with step annotations. A reward model uses spatiotemporal token pruning to judge task completion from the instruction and execution keyframes.
This framework learns an agentic-Q model to estimate individual actions’ contributions to GUI task completion. It then optimizes a policy with step-level samples from its own trajectories, decoupling policy updates from environment interaction.
ADMIRE extracts milestones from successful GUI explorations to provide adaptive intermediate rewards. It assigns credit differently to successful and failed trajectories and evaluates training on AndroidWorld, with additional transfer studies in web and embodied settings.
ACuRL uses autonomous interface exploration and an adaptive task curriculum to continually train computer-use policies with reinforcement learning. CUAJudge supplies outcome feedback, and experiments study adaptation across desktop applications and environment changes.
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.
WebArbiter generates reasoned action-preference verdicts for web navigation, using reasoning distillation followed by reinforcement learning. It introduces WebPRMBench and evaluates both preference judgments and reward-guided trajectory search.
GAIA trains a GUI action critic on positive and negative examples, then uses that critic to collect improved examples for another training round. It studies how recycling these data affects the test-time action choices of different GUI agents.
MagicGUI-RMS combines domain-specific and general reward models to assess GUI actions and propose corrections. A structured data pipeline and feedback-recycling process support repeated reward-model and agent improvement.
EchoTrail-GUI builds a memory of successful GUI trajectories through autonomous exploration filtered by a reward model. It retrieves relevant trajectories as in-context guidance for new tasks and evaluates the approach on AndroidWorld and AndroidLab.
OS-Oracle synthesizes GUI critic examples and trains a pre-action critic with supervised learning followed by consistency-preserving GRPO. It introduces OS-Critic Bench across mobile, web, and desktop settings and tests critic-assisted agents in OSWorld and AndroidWorld.
Step-GUI uses calibrated trajectory rewards to turn generated GUI interaction data into training signals for compact agent models. The report also introduces GUI-MCP for hierarchical GUI operations and local task delegation, and AndroidDaily for static-action and end-to-end mobile evaluation.
Introduces a pipeline that filters individual steps from noisy computer-agent rollouts and augments them with reasoning for training. WebSTAR and WebSCORE support GUI-policy supervision and a distilled process reward model, StepRM.
Introduces Behavior Judge, which converts execution trajectories into behavior narratives and compares multiple candidate rollouts. Experiments examine this selection approach on OSWorld, WindowsAgentArena and AndroidWorld.
GUI-PRA evaluates GUI progress by generating task-specific criteria and actively inspecting visual evidence with tools. It tests offline judgment on OS-Critic Bench and the effects of this feedback on online mobile policies in AndroidWorld.
ProRe evaluates GUI task outcomes by having a reasoner assign targeted environment-probing tasks to specialist actors. The additional observations support reward judgments, with separate evaluation of judge accuracy and downstream agent task success.
Orcust combines principle-constrained reward modeling with GUI trajectory collection in instrumented virtual machines. Environment-checkable and model-derived principles guide stepwise feedback for grounding, basic operations, and complete task execution.
TGPO merges semantically equivalent web states into tree-structured trajectories for offline preference optimization. A process reward model scores subgoal progress, redundant behavior, and action validity, while dynamic weighting emphasizes consequential decisions.
CRAFT-GUI organizes GUI trajectories by difficulty for curriculum-based GRPO training. Its reward combines rule-based signals with model judgments, with evaluation on Android Control and an internal online benchmark.
GUI-G2 replaces binary grounding rewards with Gaussian point rewards and spatial-coverage rewards. An adaptive variance tied to element dimensions supplies continuous training feedback for targets of different sizes, evaluated on the ScreenSpot benchmark family.
Introduces WebGuard, a human-annotated dataset for predicting outcomes and risk levels of state-changing website actions. It provides generalization splits and evaluates specialized guardrail models.
Guides visual GUI agents at inference time with a process-reward model that evaluates intermediate actions. It combines stepwise guidance with trajectory reflection and retry mechanisms in static and interactive navigation tasks.
Introduces StepAgent, which compares agent and expert actions to supply intermediate rewards for policy learning, with implicit-reward and inverse-RL variants. Evaluation includes WebShop and Mind2Web alongside non-web tasks.
Compares automatic trajectory evaluators for web navigation and device control against oracle judgments. The resulting feedback is used for agent fine-tuning and inference-time guidance, including experiments on WebArena.
A benchmark and evaluation toolkit for comparing automatic web-agent judges against expert judgments of task success, side effects and repetition.
Human-annotated browser-agent trajectories for comparing automated task-success and process-quality verifiers against reviewer judgments.
A cross-platform GUI-critic benchmark with mobile, desktop and web JSONL splits and screenshot assets, accompanied by public inference and accuracy-scoring scripts.
An offline preference-evaluation set for web process reward models, pairing page and trajectory context with preferred and rejected candidate decisions.
Mobile GUI reward annotations released for training UI-Genie’s trajectory-assessment model.
Web-task context, starting observations and action history paired with structured checklists and subgoal analyses for checklist generation.
Checklist-annotated web decision records with page observations, prior actions, one preferred continuation and rejected alternatives for process-reward learning.
A Qwen2.5-VL grounding checkpoint trained with Gaussian point and coverage rewards whose granularity adapts to interface-element scale.
A visual trajectory judge that retains selected intermediate screenshots to evaluate whether a web agent completed its task.
The outcome-supervision checkpoint used by WebRL to score browser trajectories during policy training.
A public text-based process-reward checkpoint from the Web-Shepherd collection for scoring web-agent decisions against task checklists.
Training and evaluation code for a mobile-agent framework that iteratively generates trajectories using agent and reward-model feedback.
Training code that learns a value environment model from offline Android interaction data and uses its frozen value estimates to guide GUI policy optimization.
Process-reward inference and benchmark-evaluation utilities that use task checklists to assess individual web-agent decisions.
Get full records and evidence through the topic export manifest. Source records can be passed to the standard RAG and citation exporters after cloning the catalog.
