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

WebFactory: Automated Compression of Foundational Language Intelligence into Grounded Web Agents

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Agentic benchmarks and orchestration · 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 WebFactory, an automated closed-loop RL pipeline that synthesizes offline websites, generates knowledge-aware tasks and LLM-collected trajectories, and trains GUI web agents (Qwen2.5-VL-3B) with a decomposed reward and GRPO.

Problem

Training GUI/web agents relies on either unsafe, non-reproducible live-web interaction or costly human-annotated data and hand-built environments.

Approach

Use LLMs to automatically synthesize versioned offline websites, generate knowledge-grounded tasks with verifiable answers, collect trajectories from a strong computer-use executor, and train a small VLM with GRPO using a decomposed format + accuracy reward (action type, click point, text F1), framed as "intelligence compression" with an LLM-embodiment comparison across foundation models.

Key results

• WebFactory-3B reaches 71.8% TCR on operational offline tasks vs 68.2% for GUI-R1-3B and 18.3% for QwenVL2.5-3B.
• Offline-to-online average TCR of 53.4% vs 37.0% for GUI-R1-3B and 20.4% for QwenVL2.5-3B.
• On GUI-Odyssey, WebFactory-3B reaches 66.0% type accuracy vs 54.8% for GUI-R1-3B.

Surprising findings

• Agent trained only on synthetic data from 10 websites matches or slightly exceeds GUI-R1-3B, which was trained on human-annotated data.

Limitations

• No exhaustive ablation of reward mechanisms against sparser or LLM-generated rewards.
• Generalization to other GUI paradigms (games, creative software) not validated.
• Offline-to-online evaluation uses only 90 tasks, and the reported improvements are based on small samples.

Experimental setup

Datasets: WebFactory synthetic offline website trajectories GUI-R1 training data

Benchmarks: WebFactory internal offline website benchmark Amazon/Airbnb/Booking offline-to-online transfer GUI-Act-Web OmniAct-Desktop GUI-Odyssey ScreenSpot

Metrics: task completion rate accuracy F1 efficiency type accuracy grounding accuracy success rate

Baselines: QwenVL2.5-3B GPT-4o GUI-R1-3B

Builds on: GUI-R1 GRPO Qwen2.5-VL WebArena BrowserGym

Closest papers by meaning

• WebRL: Training LLM Web Agents via Self-Evolving Online Curriculum Reinforcement Learning (ICLR 2025 · similarity 0.91)
• How to Train Your LLM Web Agent: A Statistical Diagnosis (NeurIPS 2025 · similarity 0.91)
• WebGym: Scaling Training Environments for Long-Horizon Visual Web Agents with Realistic Tasks (CVPR 2026 · similarity 0.90)
• A Subgoal-driven RL Framework for Improving Long-Horizon Web Agents (NeurIPS 2026 · similarity 0.90)
• WebArena: A Realistic Web Environment for Building Autonomous Agents (ICLR 2024 · similarity 0.90)
• A Real-World WebAgent with Planning, Long Context Understanding, and Program Synthesis (ICLR 2024 · similarity 0.89)