Research map / Multi-agent systems and games / LLM agents

AutoWebWorld: Synthesizing Infinite Verifiable Web Environments via Finite State Machines

ICML 2026 · Multi-agent systems and games › LLM agents · approach: GUI agent memory and trajectories · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes AutoWebWorld, which models synthetic websites as finite state machines so BFS over the FSM yields intrinsically verified GUI trajectories, and shows that training a Qwen2.5-VL-7B agent on 16K synthetic steps with GRPO reaches 27.42% on WebVoyager.

Problem

Real-website GUI trajectories are expensive to collect and hard to verify, because the underlying web state is hidden and step-level correctness depends on inconsistent external human or LLM judges.

Approach

Generate each website from an explicit FSM with pages, signature variables, preconditions and deterministic effects, then enumerate goal-reaching trajectories by BFS, replay them in the generated site with Playwright, and keep only executable ones, so success is certified by reaching a goal state rather than by an external judge.

Key results

• Ours-7B (Qwen2.5-VL-7B trained with AutoWebWorld data) reaches 27.42% overall on WebVoyager within 15 steps, the best among open-source baselines, versus 26.51% for UI-TARS-1.5-7B and 5.62% for base Qwen2.5-VL-7B.
• Success on WebVoyager and Online-Mind2Web rises monotonically with training size (8, 256, 1,024, 16,253 samples): WebVoyager 3.92% to 27.42%, Online-Mind2Web 1.22% to 14.02%.
• On ScreenSpot-Pro, Ours-7B improves from 23.2 to 27.5 and Ours-3B from 13.3 to 18.0; on ScreenSpot-V2, Ours-7B improves from 84.83 to 86.16.

Surprising findings

• Ours-3B (15.09% on WebVoyager) surpasses several 7B-plus baselines such as UI-Venus-7B and TongUI-7B despite training on only 16K steps.
• On the authors' 25-trajectory synthetic benchmark, both UI-TARS-1.5-7B and Claude-4-Sonnet score lower than on WebVoyager, which the authors read as the synthetic sites being no easier than real ones.

Limitations

• WebVoyager evaluation uses only 9 stable websites after filtering out sites with access-denied errors or CAPTCHAs, and the judge is an LLM (Gemini-3-Flash).
• Scaling analysis fits only four data points with a simple polynomial fit, so the claimed scaling law is weakly supported.
• Synthetic sites may differ from real websites; the 25-trajectory benchmark is small and built from only two cloned sites.

Experimental setup

Datasets: AutoWebWorld synthetic trajectories ScreenSpot-V2 ScreenSpot-Pro

Benchmarks: WebVoyager Online-Mind2Web ScreenSpot-V2 ScreenSpot-Pro AutoWebWorld

Metrics: task success rate grounding accuracy coordinate reward

Baselines: GPT-5.1 Claude-4-Sonnet Gemini3-Pro Qwen2.5-VL-3B Qwen2.5-VL-7B Qwen2.5-VL-72B-Instruct UI-Venus-7B OpenCUA-7B TongUI-7B UI-TARS-1.5-7B

Builds on: GRPO Playwright WebVoyager Mind2Web

Closest papers by meaning

• WebWorld: A Large-Scale World Model for Web Agent Training (ICML 2026 · similarity 0.89)
• REAL: Benchmarking Autonomous Agents on Deterministic Simulations of Real Websites (NeurIPS 2025 · similarity 0.89)
• A Real-World WebAgent with Planning, Long Context Understanding, and Program Synthesis (ICLR 2024 · similarity 0.87)
• TimeWarp: Evaluating Web Agents by Revisiting the Past (NeurIPS 2026 · similarity 0.87)
• RealWebAssist: A Benchmark for Long-Horizon Web Assistance with Real-World Users (AAAI 2026 · similarity 0.87)
• WebFactory: Automated Compression of Foundational Language Intelligence into Grounded Web Agents (ICLR 2026 · similarity 0.87)