Annotation-Free GUI Agent Data Flywheel: MobileForge’s Self-Exploring, Self-Feedback Loop

Large‑model‑driven mobile GUI agents can understand screens, click buttons, and input text, but adapting them to the rapidly evolving, fragmented app ecosystem traditionally requires costly manual task authoring, expert demonstrations, and reward labeling.

Researchers from Zhejiang University APRIL Lab, Kuaishou’s main‑site tech team, and Tsinghua University propose MobileForge, an annotation‑free, self‑exploring, self‑feedback, self‑optimizing system that closes this loop.

The paper, titled MobileForge: Annotation‑Free Adaptation for Mobile GUI Agents with Hierarchical Feedback‑Guided Policy Optimization , summarizes the core idea: let the agent explore target apps, automatically generate executable tasks, evaluate its execution with fine‑grained hierarchical feedback, and convert that feedback into trainable policy‑optimization signals.

In experiments, using only automatically generated annotation‑free data, MobileForge improves the general‑purpose vision‑language model Qwen3‑VL‑8B on the AndroidWorld benchmark from 55.2% to 67.2% Pass@3, approaching the closed‑source GUI‑specific model GUI‑Owl‑1.5‑8B (69.0%). After adapting GUI‑Owl‑1.5‑8B, the derived ForgeOwl‑8B reaches 77.6% Pass@3 on AndroidWorld and 41.0% success on the out‑of‑domain MobileWorld GUI‑only tasks.

MobileForge consists of two coupled components:

MobileGym : an interaction and evaluation backbone that explores reachable states in the target app, generates executable tasks from the exploration evidence, and provides fine‑grained, hierarchical evaluation of the agent’s rollouts.

HiFPO (Hierarchical Feedback‑Guided Policy Optimization): schedules multiple attempts, reuses failure hints, filters tasks, and updates the model with hint‑contextualized step‑level GRPO.

The overall pipeline is: target‑app exploration → curriculum generation → multiple rollouts → hierarchical evaluation → task/trajectory/step filtering → hint‑contextualized GRPO training. No manual tasks, expert demos, or human‑labeled rewards are involved.

MobileGym has three stages:

Target‑App Exploration : the agent enters the app, uses APK activity information and screenshots to set functional exploration goals, and performs a depth‑first‑like traversal, recording state transitions, actions, target elements, metadata, and natural‑language summaries.

MobileGym‑Curriculum : each exploration trajectory is assessed for coherence and goal completion, then multiple task variants are generated. A task is represented as a five‑tuple (instruction, step‑budget, core function, variation type, preconditions), anchored to observed app behavior.

MobileGym‑Critic : a hierarchical evaluator produces three feedback types for each rollout: trajectory‑level outcome label, step‑level process label, and a corrective hint summarizing failure reasons and suggested alternatives.

HiFPO turns the hierarchical feedback into training signals:

Hint‑augmented multiple attempts : for each task the current policy tries K times; after a failure, the critic’s hint is appended to the task instruction for the next attempt, allowing the agent to accumulate experience.

Task filtering : tasks that are already mastered (all attempts succeed) are removed; tasks that are fully failed or partially successful are retained for learning.

Trajectory and step selection : among retained tasks, the highest‑quality successful trajectory is chosen; if all fail, the trajectory with the highest proportion of reasonable steps is selected, and only those steps are kept for training.

Hint‑contextualized step‑level GRPO : each step‑level sample includes the task, screenshots, interaction history, and the hint; the model samples multiple candidate actions and compares them using a regularized GUI‑action reward.

Experimental setup uses two benchmarks: AndroidWorld (in‑domain) with 116 tasks and MobileWorld GUI‑only (cross‑domain) with 117 tasks. Base models are Qwen3‑VL‑8B (general VLM) and GUI‑Owl‑1.5‑8B (GUI‑specific). MobileForge generates 3,249 candidate tasks from 20 apps (527 source trajectories) and trains on subsets of 200, 400, and 900 tasks.

Results:

Qwen3‑VL‑8B baseline Pass@3: 55.2%; after 900 auto‑generated tasks, ForgeQwen3‑8B reaches 67.2% Pass@3 (Pass@1: 40.5%→50.9%, Pass@2: 49.1%→60.3%).

GUI‑Owl‑1.5‑8B baseline Pass@3: 69.0%; after adaptation, ForgeOwl‑8B achieves 77.6% Pass@3 (Pass@1: 56.0%→67.2%).

Cross‑domain MobileWorld success: ForgeOwl‑8B 41.0% vs. GUI‑Owl‑1.5‑8B 37.6%; ForgeQwen3‑8B improves from 7.6% to 10.3%.

Several ablation studies isolate the contributions of hint usage, training objectives, task filtering, critic model choice, and curriculum grounding. For example, adding correction hints raises total success from 52.0% to 77.0% and Pass@3 from 49.0% to 72.5%; hint‑contextualized GRPO consistently outperforms no‑hint SFT and hint‑SFT across task counts.

A case study on the Pro Expense app shows that the baseline Qwen3‑VL‑8B loses task intent after early deletions, whereas ForgeQwen3‑8B maintains the deletion pattern and completes all three expense removals.

Tag‑wise failure‑rate analysis reveals strong improvements on verification, search, complex UI, screen reading, repetition, and information‑retrieval tasks, while game‑playing, multi‑app, memorization, and math‑counting remain challenging.

In conclusion, MobileForge provides a fully annotation‑free adaptation loop for mobile GUI agents: MobileGym turns real‑app interactions into tasks and hierarchical feedback, and HiFPO converts that feedback into step‑level policy updates, enabling continuous self‑improvement without human‑written data.