PaperlayerView exact revision on arXiv

Source-linked AI summary

AutoWebWorld: Synthesizing Infinite Verifiable Web Environments via Finite State Machines

Yifan Wu, Yiran Peng, Yiyu Chen, Jianhao Ruan, Zijie Zhuang, Cheng Yang, Jiayi Zhang, Man Chen, Yenchi Tseng, Zhaoyang Yu, Liang Chen, Yuyao Zhai, Bang Liu, Chenglin Wu, Yuyu Luo

arXiv:2602.14296v1cs.AIcs.SE

TL;DR

Autonomous Web GUI agents lack sufficient high-quality, verifiable trajectories because real websites hide state transitions and require costly external judging. AutoWebWorld models environments as FSMs, generates executable sites, and verifies trajectories through explicit rules and goal states. It produces 11,663 verified trajectories across 29 environments at $0.04 each, while synthetic-data training improves real-world benchmark performance and scales consistently with data volume.

  • Problem

    High-quality, verifiable web interaction trajectories are scarce because real websites expose observations while hiding internal state transitions, requiring costly and inconsistent external verification.

  • Method

    AutoWebWorld generates FSM specifications with explicit states, actions, and transition rules, translates them into interactive websites, and searches and verifies trajectories programmatically.

  • Results

    11,663 verified trajectories across 29 environments were generated at $0.04 per trajectory, while a 7B agent trained on 16K steps achieved 27.42% on WebVoyager and performance improved consistently with more synthetic data.

  • Takeaways & Limitations

    Verified synthetic web environments support reproducible trajectory synthesis and transfer to real-world web benchmarks, with performance improving as synthetic data scales.

  • Takeaways & Limitations

    The framework distinguishes internal state from potentially ambiguous visual observations, which previous screenshot-based approaches commonly use to represent state.

Abstract

from arXiv · show

The performance of autonomous Web GUI agents heavily relies on the quality and quantity of their training data. However, a fundamental bottleneck persists: collecting interaction trajectories from real-world websites is expensive and difficult to verify. The underlying state transitions are hidden, leading to reliance on inconsistent and costly external verifiers to evaluate step-level correctness. To address this, we propose AutoWebWorld, a novel framework for synthesizing controllable and verifiable web environments by modeling them as Finite State Machines (FSMs) and use coding agents to translate FSMs into interactive websites. Unlike real websites, where state transitions are implicit, AutoWebWorld explicitly defines all states, actions, and transition rules. This enables programmatic verification: action correctness is checked against predefined rules, and task success is confirmed by reaching a goal state in the FSM graph. AutoWebWorld enables a fully automated search-and-verify pipeline, generating over 11,663 verified trajectories from 29 diverse web environments at only $0.04 per trajectory. Training on this synthetic data significantly boosts real-world performance. Our 7B Web GUI agent outperforms all baselines within 15 steps on WebVoyager. Furthermore, we observe a clear scaling law: as the synthetic data volume increases, performance on WebVoyager and Online-Mind2Web consistently improves.

1. Introduction

AutoWebWorld addresses the scarcity and unverifiability of web interaction trajectories by making environment state transitions explicit and searchable. Its synthetic data and environments yield verified trajectories, strong WebVoyager performance, and consistent gains as synthetic data scales.

  • High-quality, verifiable interaction trajectories remain scarce, limiting training data for autonomous Web GUI agents.
  • Observation-based collection cannot directly check internal state transitions, so existing methods rely on inconsistent and costly external verifiers.
  • AutoWebWorld generates FSMs with explicit states, actions, and transition rules, then uses BFS over the known graph to collect trajectories systematically.
  • 11,663 reproducibly verified trajectories were synthesized across 29 interactive web environments at $0.04 per trajectory.
  • 27.42% is achieved on WebVoyager by a 7B agent trained on 16K synthetic steps, while performance consistently improves as synthetic data increases.
  • The framework contributes an FSM-based paradigm and releases diverse environments and verified trajectories as training data and a reproducible benchmark.

2. Related Work

Prior web trajectory collection relies mainly on real environments, where externally verified correctness makes scaling difficult. AutoWebWorld extends synthetic-environment work with transition-driven generation, intrinsic verification, and a four-step search-and-filter pipeline.

  • Prior approaches collect trajectories through online exploration, human demonstrations, and other real-environment methods.
  • Real-website trajectory scaling is difficult because correctness must be determined by external verifiers.
  • AutoWebWorld uses a four-step pipeline: FSM generation, website synthesis, BFS trajectory exploration, and Playwright-based replay filtering.
  • AutoWebWorld advances synthetic web environments by explicitly modeling known state-transition structure while supporting reproducible benchmarking and batch data synthesis.

3. Preliminary

The paper formalizes web GUI interaction as deterministic state evolution in an MDP, distinguishing internal semantic state from visual observation. States combine page identity with structured signatures, while preconditions and effects determine transitions.

  • GUI Trajectory: Web GUI interaction is modeled as an MDP with internal states, executable actions, deterministic transitions, and observations.
  • GUI Trajectory: The agent’s visual observation is distinct from the internal state used for planning, trajectory synthesis, and automatic verification.
  • State Definition: Each internal state is represented by a page identity and a page signature containing structured variables for page configuration and task context.
  • State Definition: Signature variables can encode query text, filters, sorting, pagination, form values, selected items, and cart contents.
  • State Transition: The transition function maps a state and action to the next state, with applicability checked by preconditions.
  • State Transition: Invalid actions follow a fixed rule, such as a no-op, while valid actions apply deterministic effects to update the state signature.
  • State Transition: Navigation initializes the target page signature and merges carried variables, making the next state uniquely determined and searchable.

4. AutoWebWorld

AutoWebWorld constructs executable web environments from validated FSMs, then searches their state graphs for grounded trajectories and filters them through front-end execution. The resulting dataset combines intrinsic verification, long horizons, and low-cost reproducibility.

  • FSM Generation: AutoWebWorld generates FSM specifications defining pages, state signatures, executable transitions, and goal states for each synthesized website.A validator checks reachability of terminal states and alignment between page-transition logic and expected web behavior.
  • Web Environment Generation: Coding agents translate validated FSMs into runnable simulated websites whose interactive behavior follows the FSM semantics.The pipeline generates guidelines, synthesizes pages with iterative review, builds the project, and repairs build failures when necessary.
  • Trajectory Synthesis: Breadth-first search traverses FSM state graphs, mapping each high-level action to a predefined sequence of atomic GUI operations that realizes the transition.Executing these procedures produces grounded trajectories aligned with FSM paths.
  • Automatic Trajectory Filtering: Execution-based filtering accepts a candidate only when every atomic GUI operation succeeds during sequential front-end replay.This strict rule removes trajectories that are valid at the FSM level but fail because of implementation mismatches.
  • Datasets Analysis: 11,663 verified trajectories across 29 environments cost $0.04 per trajectory, while the average trajectory length reaches 21.94 steps.The synthesized environments can also serve as stable GUI benchmarks with consistent success criteria.

5. Experiments

Experiments evaluate navigation, grounding transfer, scaling, and benchmark stability. AutoWebWorld data improves real-world GUI-agent performance, scales predictably with data volume, and supports challenging reproducible evaluations.

  • Navigation Results on WebVoyager: AutoWebWorld trajectories improve real-world navigation, with Ours-7B achieving 27.42% on WebVoyager and outperforming the listed open-source baselines.Evaluation uses a 15-action limit and reports results by website domain.
  • Navigation Results on WebVoyager: 16K synthesized training steps enable Ours-3B to reach 15.09% overall on WebVoyager, surpassing several larger-data 7B+ baselines.The comparison includes UI-Venus-7B trained with roughly 1M SFT samples and TongUI-7B with about 350K RL samples.
  • Grounding Results: AutoWebWorld-derived grounding data consistently improves both ScreenSpot-V2 and ScreenSpot-Pro performance for the evaluated 3B and 7B models.On ScreenSpot-Pro, Ours-3B rises from 13.3 to 18.0 and Ours-7B from 23.2 to 27.5.
  • Importance of Grounding Data: After roughly 25 training steps, including grounding data produces faster, more stable coordinate-reward growth and a higher sustained total-reward plateau.Without grounding, the initial total reward is briefly higher because harder training instances are excluded.
  • Benchmark Stability: AutoWebWorld benchmark success rates are lower than WebVoyager for both evaluated agents, indicating that synthesized sites are not trivially easier.The benchmark remains controllable and reproducibly testable while preserving comparable or stronger interaction challenges.

6. Conclusion

AutoWebWorld uses FSM specifications to synthesize controllable web environments and make large-scale search-and-verify trajectory generation practical. Across 29 environments, it produces verified trajectories cheaply, while training shows strong real-world benchmark gains and consistent improvement with more synthetic data.

  • AutoWebWorld synthesizes web environments from FSM specifications, enabling practical and low-cost large-scale search-and-verify trajectory generation.Different FSM page sets and transition structures also provide controllable difficulty.
  • 29 environments yield 11,663 verified trajectories at $0.04 per trajectory.
  • Training with only 16k GRPO steps yields strong gains on real-world benchmarks.
  • Performance on real-world benchmarks consistently improves as synthetic data scales.

Impact Statements

AutoWebWorld makes web interaction states and success criteria explicit, enabling intrinsic verification and reproducible trajectory synthesis. Its FSM design also supports systematic control of environment difficulty and state-space complexity.

  • AutoWebWorld verifies task success inside the environment instead of relying on external judges.This reduces ambiguity in web GUI trajectory collection.
  • FSM design enables systematic control of environment difficulty through different page sets and transition structures.
  • Each environment uses fsm.json components for metadata, pages, actions, and a navigation skeleton.
  • Terminal pages define success, with the simplest intrinsic criterion being reachability of a terminal page.
  • Terminal pages must be reachable from the initial page and correspond to unambiguous goal attainment.
  • Internal states pair a page id with a task-relevant signature, and BFS deduplicates states using (p, hash(σ)).Signatures make otherwise implicit semantic state explicit while avoiding unnecessary UI noise.

A.1.3. ACTION SCHEMA AND DETERMINISTIC TRANSITION SEMANTICS

AutoWebWorld defines executable FSM actions through explicit preconditions, deterministic effects, and optional navigation, separating semantic outcomes from GUI execution procedures. These rules make successor states uniquely determined and support reproducible BFS-based trajectory synthesis.

  • Each action specifies preconditions, effects, navigation status, an optional target page, and a GUI procedure.
  • Preconditions are conjunctions evaluated only on the current signature, not screenshot-level element existence.This makes action availability verifiable and robust to rendering variation.
  • Effects apply only when all preconditions hold and update only explicitly declared signature fields deterministically.
  • Navigation applies effects before switching pages, then initializes and merges the target page signature deterministically.
  • Result-set-changing actions must reset pagination fields to prevent stale page-index semantics.
  • Given (p, σ, a), every executable action has a uniquely determined successor state, enabling step-wise verification.
  • GUI procedures expand semantic actions into GUI-atomic operations, while semantic correctness remains defined by signature-space preconditions and effects.
  • BFS uses executable actions, signature-based deduplication, and bounded expansion to enumerate reachable trajectories reproducibly.Navigation skeletons provide structural connectivity checks but do not add transition semantics.

A.2.3. MAPPING TASKS TO GOAL PREDICATES IN SIGNATURE SPACE

AutoWebWorld maps tasks to deterministic goal predicates over explicit semantic states. BFS identifies goal-reaching paths, and replay verifies each transition and final goal satisfaction intrinsically.

  • Tasks are represented as goal predicates G(s) over semantic states, with terminal-page reachability as the simplest form.
  • Goal predicates can combine signature constraints such as selected items, filled fields, or cart contents.Because constraints use explicit signature paths, evaluation does not depend on screenshot heuristics.
  • BFS checks G(s) when dequeuing states, and each satisfying path is shortest among trajectories reaching that goal key.
  • Replay rechecks preconditions, effects, navigation semantics, and final goal satisfaction for every discovered trajectory.
  • Distinct trajectories can reach the same goal through different semantic keys, while truncated or invalid trajectories support robustness training.

A.2.5. LINKING SEMANTIC ACTIONS TO EXECUTABLE GUI PROCEDURES

AutoWebWorld links FSM-level actions to executable GUI procedures, preserving a semantic representation for state evolution and verification alongside atomic operations for replay and supervision.

  • Each BFS semantic action receives an attached GUI procedure that forms an executable GUI-atomic sequence.This creates a two-level representation separating semantic state transitions from concrete interface operations.
  • Semantic trajectories support state evolution and verification, while GUI-level trajectories support replay and coordinate-level supervision.
  • Correctness is determined by the FSM-defined state update and goal predicate rather than by the GUI procedure itself.The GUI procedure supplies an execution realization, while semantic transitions remain the source of truth.
  • The FSM specification organizes environments around pages, actions, preconditions, effects, and grounded GUI procedures.Actions may encode navigation, state assignments, search, sorting, or item selection through executable interface operations.
  • The specification also records initial pages, terminal pages, page signatures, available actions, and cross-page connectivity.

B. Datasets Construction Details and Statistical Analysis

The dataset construction details describe environment-side data files and provide representative records for providers, therapists, prescriptions, and appointments.

  • Each synthesized website is converted into query instances with intrinsic ground truth and reproducible statistical-analysis artifacts.The pipeline assumes environment-side files including fsm.json and bfs.json, alongside data.js examples.
  • The example data include provider records with identifiers, names, specialties, ratings, images, and next available slots.
  • Therapist records encode identifiers, names, specialties, experience values, and image paths.
  • Prescription records specify medication names, dosages, statuses, supplies, and image paths.
  • Appointment records include providers, dates, times, visit types, statuses, and image paths.

C. Cost Analysis

The cost analysis decomposes AutoWebWorld construction expenses into website generation, FSM production, query generation, and reasoning-trace generation.

  • The pipeline cost is divided into Web, FSM, Queries, and Thinking categories.These represent coding-agent website generation, multi-agent FSM specification, language-model query generation, and step-level reasoning traces.

D. Model Implementation and Training Details

The implementation uses distributed GRPO training with structured rewards for action type, coordinate grounding, and output format, supported by a unified GUI action space.

  • Training and inference run on a single machine with 8× NVIDIA A800 GPUs using distributed PyTorch training.
  • The policy initializes from Qwen2.5-VL-7B and uses a 1024-token maximum prompt length.
  • The effective batch size is 256 samples per optimization step, with BF16 precision, gradient checkpointing, FlashAttention-2, and ZeRO-3 sharding.The batch size combines per-device batch size 4, 8 accumulation steps, and 8 GPUs.
  • The training setup runs for 1 epoch and generates 8 rollouts per prompt.
  • The composite GRPO reward combines action-type accuracy, coordinate grounding, and format compliance.
  • Coordinate grounding checks whether click or hover predictions fall inside the ground-truth bounding box, while other actions use action-type correctness.
  • Format reward requires completions to follow the ordered, closed-tag template <think>...</think><action>...</action>.
Loading 2602.14296v1…