AutoWebWorld: Synthesizing Infinite Verifiable Web Environments via Finite State Machines

AutoWebWorld: Synthesizing Infinite Verifiable Web Environments via Finite State Machines

Overview

AutoWebWorld is a framework that synthesizes controllable and verifiable web environments by modeling them as Finite State Machines (FSMs). This enables programmatic generation and verification of large-scale, high-quality GUI interaction datasets, in contrast to the limitations of existing real-world data collection pipelines.

Problem & Context

• The performance of autonomous Web GUI agents heavily relies on the quality and quantity of training data.
• Collecting high-quality interaction trajectories from real websites is expensive and difficult to verify, as the underlying state transitions are hidden from the agent.
• Existing data collection methods rely on external verifiers (human annotators or LLM judges), leading to inconsistency and high cost.

Methodology

1. FSM Generation: AutoWebWorld generates an FSM specification for each website, which explicitly defines all states, actions, and transition rules.
2. Web Environment Generation: The FSM is translated into a runnable front-end website using coding agents, enabling deterministic replay and verification of GUI interactions.
3. Automatic Trajectory Collection: Breadth-first search is used to systematically explore the FSM state graph and enumerate goal-reaching trajectories, which are then verified by executing them in the generated websites.

Data & Experimental Setup

• AutoWebWorld synthesized 29 diverse web environments and over 11,663 verified trajectories across these environments.
• The synthesized data is used to train a 7B-parameter Web GUI agent, which is evaluated on the WebVoyager benchmark.

Results

• The 7B-parameter agent trained on 16K steps of AutoWebWorld data achieves a 27.42% success rate on WebVoyager, outperforming all baselines.
• As the amount of synthesized data increases, the agent's performance on WebVoyager and Online-Mind2Web consistently improves, demonstrating the scalability potential of this approach.

Interpretation

• AutoWebWorld's state-driven paradigm enables scalable, verifiable data synthesis, overcoming the "verifier bottleneck" that limits real-world data collection.
• The verified synthetic data allows training highly capable Web GUI agents, demonstrating the value of intrinsic environment verification over relying on external judges.
• The clear scaling trend suggests that further increasing the volume of synthesized data could lead to even stronger real-world performance.

Limitations & Uncertainties

• While the synthesized environments capture essential web interaction patterns, they do not fully reflect the visual complexity and unpredictability of real websites.
• The current AutoWebWorld pipeline requires significant engineering effort to specify the FSM for each environment. Automating this process could further improve scalability.

What Comes Next

• Exploring techniques to automatically generate FSM specifications from real website samples, reducing the manual effort required.
• Investigating ways to incorporate more realistic rendering and dynamic website behaviors into the synthetic environments.
• Studying the generalization capabilities of agents trained on AutoWebWorld data to even broader web interaction tasks beyond the current benchmarks.

Sources:

• AutoWebWorld: Synthesizing Infinite Verifiable Web Environments via Finite State Machines