CRAFT-GUI: Curriculum-Reinforced Agent for GUI Tasks
- CRAFT-GUI is a curriculum-guided reinforcement learning framework that automates mobile GUI tasks by decomposing rewards into detailed subcomponents.
- It employs stage-wise training with difficulty modeled by trajectory length, enabling stable optimization across easy to complex tasks.
- Empirical results show CRAFT-GUI outperforms baselines, achieving up to a 10.3% improvement on challenging mobile GUI benchmarks.
CRAFT-GUI, short for Curriculum-Reinforced Agent For GUI Tasks, is a curriculum-guided reinforcement learning framework for GUI task automation built on Group Relative Policy Optimization (GRPO). It is designed to address two limitations identified in prior RL-based GUI agents: treating training data as a uniform set despite substantial variation in GUI-task difficulty, and relying on coarse rewards that collapse tool choice, argument correctness, formatting, and semantic adequacy into a single signal. CRAFT-GUI therefore combines stage-wise curriculum learning with fine-grained verified rewards, and is evaluated on AndroidControl, AndroidWorld, and an internal online benchmark derived from real human interaction logs (Nong et al., 15 Aug 2025).
1. Research setting and problem formulation
CRAFT-GUI is situated in the literature on autonomous agents that operate graphical user interfaces by perceiving screens and issuing actions such as clicking, typing, swiping, or invoking system controls. Its starting observation is that GUI tasks are not uniformly difficult. Some tasks are short and atomic, while others require long trajectories, cross-app navigation, or visual-semantic interpretation. The framework treats this heterogeneity as a first-class optimization problem rather than as noise in the training distribution.
The paper identifies two principal deficiencies in earlier RL-based GUI-agent training. First, optimization becomes unstable when easy and hard tasks are mixed without structure: difficult trajectories can dominate gradients early, while easy trajectories become insufficiently informative later. Second, a single binary success signal is too coarse for GUI automation, because correctness depends on several distinct subdecisions, including the selected tool, the arguments or coordinates, response format, semantic adequacy for understanding tasks, and response length. CRAFT-GUI’s central thesis is that curriculum learning plus richer verified rewards yields more stable and more effective policy optimization for GUI interaction (Nong et al., 15 Aug 2025).
The framework is primarily developed in a mobile GUI setting, but the problem statement is broader. The paper explicitly suggests future extension to computer usage scenarios and to trial-and-error with rollback mechanisms, indicating that the proposed training principles are intended to generalize beyond the immediate benchmark suite.
2. GRPO-based framework and stage-wise curriculum
At the optimization level, CRAFT-GUI is built on GRPO rather than PPO. The stated reason is architectural economy: unlike PPO, GRPO does not require a separate value function, which reduces memory and compute overhead for large multimodal GUI agents. This choice is directly aligned with the framework’s emphasis on scalable RL over visually grounded action spaces (Nong et al., 15 Aug 2025).
The training loop combines four elements: curriculum learning, GRPO, fine-grained verified rewards, and KL regularization. The model is trained sequentially through curriculum stages. For each prompt in the current stage, the policy samples a group of candidate outputs; each candidate is scored by the reward function; the rewards are normalized within the group; and the policy is updated using the GRPO objective while a frozen reference model provides KL regularization.
The paper gives the group-relative advantage estimate as
This formulation means that a sampled trajectory is evaluated relative to peer samples produced for the same prompt, rather than against a globally calibrated scalar baseline. The intended effect is variance reduction and more stable learning.
The framework description also presents a curriculum dataset set
$\mathcal{C} = \{\mathcal{D}_{\text{basic}, \mathcal{D}_{\text{intermediate}, \mathcal{D}_{\text{advanced}, \mathcal{D}_{\text{ui}\}$
together with the update rule in which the reference policy is set from the current policy and then kept frozen during GRPO optimization. The notation is reproduced as described in the source, which also notes that the implementation follows the original GRPO advantage-estimation strategy rather than introducing a more elaborate critic.
3. Difficulty modeling and curriculum design
CRAFT-GUI defines task difficulty primarily by trajectory length, namely the number of interaction steps required to complete an instruction. This is the framework’s operational proxy for difficulty: short trajectories are treated as simpler, while longer trajectories imply more planning, stronger state dependence, and more opportunities for error. The paper also treats several visual understanding tasks as inherently harder than direct operation tasks because they require semantic interpretation rather than action execution.
| Tier | Criterion | Typical content |
|---|---|---|
| Basic | interaction steps | Simple operation tasks |
| Intermediate | $4$ to $8$ steps | Typical multi-step workflows |
| Advanced | steps | Long trajectories plus VQA, information extraction, and element localization |
Training proceeds sequentially through Stage 1: Basic, Stage 2: Intermediate, and Stage 3: Advanced (Nong et al., 15 Aug 2025). The framework presents this progression as analogous to human learning: basic UI manipulation is acquired first, followed by multi-step operations, and only then more abstract perception and reasoning tasks. In the paper’s own framing, this is a form of difficulty-aware decoupling of the training distribution.
The choice of trajectory length is explicitly acknowledged as practical rather than perfect. A short task may still be hard, and a long task may be routine. Even so, the paper treats length as a workable proxy that can be implemented consistently across large GUI-task corpora.
4. Reward decomposition and verified supervision
A major contribution of CRAFT-GUI is its hybrid reward design, which combines rule-based verification with model-judged evaluation. The objective is to replace a blunt success/failure signal with a structured reward that exposes where an action failed.
For mobile operation tasks, the overall reward is
Here, checks tool selection, checks parameter correctness, enforces structured output, and $\mathcal{C} = \{\mathcal{D}_{\text{basic}, \mathcal{D}_{\text{intermediate}, \mathcal{D}_{\text{advanced}, \mathcal{D}_{\text{ui}\}$0 penalizes overlong responses. The tool reward is defined as
$\mathcal{C} = \{\mathcal{D}_{\text{basic}, \mathcal{D}_{\text{intermediate}, \mathcal{D}_{\text{advanced}, \mathcal{D}_{\text{ui}\}$1
and the argument reward is defined as
$\mathcal{C} = \{\mathcal{D}_{\text{basic}, \mathcal{D}_{\text{intermediate}, \mathcal{D}_{\text{advanced}, \mathcal{D}_{\text{ui}\}$2
with task-specific matching logic. For click and long press, the predicted point must fall within the same widget as ground truth. For swipe, the angle between predicted and ground-truth vectors must be within $\mathcal{C} = \{\mathcal{D}_{\text{basic}, \mathcal{D}_{\text{intermediate}, \mathcal{D}_{\text{advanced}, \mathcal{D}_{\text{ui}\}$3. For type and open app, correctness is determined by semantic similarity plus character-level matching. For system button, matching is strict string equality. For wait and terminate, there is no argument reward.
The format reward is
$\mathcal{C} = \{\mathcal{D}_{\text{basic}, \mathcal{D}_{\text{intermediate}, \mathcal{D}_{\text{advanced}, \mathcal{D}_{\text{ui}\}$4
and the paper states that the response schema uses HTML-style tags such as >, <tool_call>, and <task_progress>. This makes formatting itself part of policy optimization rather than a downstream post-processing concern.
The length term is introduced to suppress “runaway reasoning-token growth.” The source describes a piecewise penalty intended to preserve useful reasoning while discouraging verbosity, though it notes formatting errors in the printed expression. The conceptual role of the term is therefore clearer than its exact typographic rendering.
For visual understanding tasks, the reward becomes
$\mathcal{C} = \{\mathcal{D}_{\text{basic}, \mathcal{D}_{\text{intermediate}, \mathcal{D}_{\text{advanced}, \mathcal{D}_{\text{ui}\}$5
where the semantic component $\mathcal{C} = \{\mathcal{D}_{\text{basic}, \mathcal{D}_{\text{intermediate}, \mathcal{D}_{\text{advanced}, \mathcal{D}_{\text{ui}\}$6 is supplied by an LLM-as-judge mechanism rather than symbolic verification (Nong et al., 15 Aug 2025). This distinction reflects the different verification regime of understanding tasks: open-ended answers often cannot be checked with exact rules.
The reward design is thus intentionally heterogeneous. It uses hard verification where exact correctness is available and model-judged feedback where the task is semantically open-ended. A plausible implication is that CRAFT-GUI treats GUI interaction not as a monolithic action problem, but as a bundle of partially separable subskills.
5. Training pipeline, benchmarks, and empirical results
CRAFT-GUI uses Qwen2.5-VL backbones, specifically 7B and 32B variants. The training data are split into basic operation data, intermediate operation data, advanced operation data, and understanding data. Policy optimization then proceeds stage by stage, with group sampling, reward computation, group-relative normalization, and a GRPO update against a frozen reference model (Nong et al., 15 Aug 2025).
The public evaluation is conducted on AndroidControl and AndroidWorld. AndroidControl reports Type, Grounding, and Success Rate (SR), and the paper defines the main metric as
$\mathcal{C} = \{\mathcal{D}_{\text{basic}, \mathcal{D}_{\text{intermediate}, \mathcal{D}_{\text{advanced}, \mathcal{D}_{\text{ui}\}$7
The experiments are evaluated under zero-shot generalization on unseen task distributions and pass@1 single-attempt execution.
The internal online benchmark is constructed from real human interaction logs and spans six application categories: food delivery, in-store dining, medical services, financial services, insurance services, and gaming. Each category contains 120–200 unique test cases, and the benchmark is intended to measure page comprehension, contextual reasoning, and decision-making.
Setting Result Public benchmarks +5.6% improvement Internal online benchmarks +10.3% improvement Internal benchmark, CRAFT-GUI-32B 75.7% average SR The abstract’s headline claim is that CRAFT-GUI outperforms previous state-of-the-art approaches by 5.6% on public benchmarks Android Control and 10.3% on internal online benchmarks (Nong et al., 15 Aug 2025). On the internal benchmark, CRAFT-GUI-32B reaches 75.7% average SR, compared with 65.4% for UI-TARS-72B, 64.8% for Qwen2.5-VL-72B, 35.9% for Claude-3.7-Sonnet, and 27.7% for GPT-4.1.
On AndroidControl and AndroidWorld, CRAFT-GUI-32B-stage3 reports 92.7% SR on AndroidControl-Low, 80.3% SR on AndroidControl-High, and 51.7% SR on AndroidWorld. The paper compares these with strong baselines including UI-TARS-72B at 91.3 / 74.7 / 46.6 and Qwen2.5-VL-72B at 93.7 / 67.4 / 35.0. The pattern emphasized in the paper is that CRAFT-GUI is competitive on lower-difficulty AndroidControl settings while stronger on harder settings and on AndroidWorld.
The training-method ablation is central to the paper’s argument. On the internal benchmark, SFT yields 60.8%, Vanilla GRPO yields 71.9%, and Curriculum GRPO yields 75.7%. The paper therefore reports +11.1% over SFT for vanilla GRPO, +14.9% over SFT for curriculum GRPO, and +3.8% over vanilla GRPO. A separate data-mixture ablation shows 73.2% for Operation only and 75.7% for Operation + Understanding, a +2.5% gain indicating that understanding data improves operation performance.
6. Interpretation, limitations, and relation to adjacent work
The paper’s interpretation of its own results is that easy-to-hard progression stabilizes optimization, richer rewards provide denser and more informative feedback, and understanding tasks strengthen visual-semantic representations that transfer back to GUI operation (Nong et al., 15 Aug 2025). In this sense, CRAFT-GUI reframes GUI-agent training from uniform RL over heterogeneous tasks into a staged regime with explicit difficulty structure and decomposed supervision.
Several limitations are also explicit. Trajectory length as difficulty proxy is practical but imperfect. The curriculum itself is hand-designed rather than adaptively learned. Rule-based reward terms depend on task structure and therefore work best for structured GUI operations. The LLM-as-judge component may introduce noise or bias. The primary empirical setting remains mobile GUI, and the length penalty is acknowledged as heuristic, with the risk of suppressing beneficial reasoning if tuned poorly.
Within adjacent GUI-agent research, CRAFT-GUI occupies a distinct niche. GUICrafter addresses GUI-agent weakness primarily as a data-scarcity problem and uses a two-stage weakly supervised curriculum built from unannotated screenshots and webpages, followed by GRPO-based calibration on a smaller high-quality annotated set (Fan et al., 29 Jun 2026). GUI-World addresses a different gap again, focusing on dynamic GUI videos, multi-window interaction, and temporal reasoning, and concludes that current multimodal models still struggle with dynamic GUI content and sparse GUI-video signals (Chen et al., 2024). Relative to these works, CRAFT-GUI’s contribution is not benchmark construction or weak supervision from raw GUI evidence, but the integration of difficulty-aware curriculum learning and fine-grained hybrid rewards into RL-based GUI interaction.
It should also be distinguished from CGui, a declarative, constraint-based language and toolchain for highly customized cross-platform GUI development, which concerns GUI construction rather than GUI-agent policy learning (Menik et al., 2023). The similarity of names can obscure the fact that the two systems belong to different problem classes: one generates or automates user interfaces, while the other trains agents to act within them.
Taken together, these distinctions place CRAFT-GUI as a reinforcement-learning framework for GUI-task execution whose primary technical claims concern curriculum structure, verified reward granularity, and empirical gains on operation-heavy benchmarks. Its broader significance lies in showing that, for GUI agents, how tasks are ordered and how correctness is decomposed can matter as much as model scale or raw data volume.