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

WebWorld: A Large-Scale World Model for Web Agent Training

ICML 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.

Trains an open-web world model (WebWorld, 8B/14B/32B Qwen3-based) on 1M+ real-web trajectories, evaluates it with a new WebWorld-Bench, and shows downstream gains for web agents and inference-time search.

Problem

Web agent training is limited by latency, rate limits, and safety risks of live browsing, while existing trained web world models rely on closed benchmark environments with about 10k trajectories and poor generalization.

Approach

A hierarchical, scalable data pipeline (randomized crawling of pre-training-corpus URLs, LLM-driven autonomous exploration, and task-oriented execution) yields 1M+ open-web transitions in A11y Tree format; a two-stage recipe (dynamics learning then a small CoT fine-tune) plus multi-format augmentation trains the simulator, which is then evaluated with the WebWorld-Bench Factuality and Web Turing s

Key results

• WebWorld-32B reaches 71.0% average Factuality on WebWorld-Bench, matching Claude-Opus-4.1 (71.3%) and close to Gemini-3-Pro (70.3%).
• Qwen3-8B fine-tuned on WebWorld-synthesized trajectories gains +9.9 points on MiniWob++ and +10.9 points on WebArena (SR 20.7%); the 14B variant reaches 24.3% on WebArena, comparable to GPT-4o's 26.6%.
• Inference-time lookahead with WebWorld and a GPT-5 value model reaches 67.5% on MiniWob versus 64.3% greedy GPT-4o baseline.

Surprising findings

• Only 1K CoT samples after dynamics training outperform Qwen3-8B direct reasoning tuning with 10x more CoT data (0.561 vs 0.510 Total Score).
• Open-source baseline scores on WebWorld-Bench are very low, which the authors attribute to output format misalignment rather than model deficiency.

Limitations

• Exhibits sycophancy, generating overly optimistic outcomes that cater to the agent's action.
• Struggles to generate high-quality detailed content such as scientific articles.
• Inference-time search yields only bounded gains; the judge for evaluation is GPT-4o, which may bias scores.

Experimental setup

Datasets: FineWeb CCI 3.0 Mind2Web AgentTrek Ultrachat WebArena MiniWob++

Benchmarks: WebWorld-Bench WebArena MiniWob++

Metrics: Factuality Score Web Turing Score success rate average steps

Baselines: GPT-4o Claude-Sonnet-4.5 Claude-Opus-4.1 Gemini-3-Pro GPT-5 WMA WebSynthesis Word2World WebEvolver DreamGym

Builds on: Qwen3 LLaMA-Factory BrowserGym UI-Simulator WebSynthesis

Closest papers by meaning

• WebGym: Scaling Training Environments for Long-Horizon Visual Web Agents with Realistic Tasks (CVPR 2026 · similarity 0.90)
• Web Agents with World Models: Learning and Leveraging Environment Dynamics in Web Navigation (ICLR 2025 · similarity 0.89)
• AutoWebWorld: Synthesizing Infinite Verifiable Web Environments via Finite State Machines (ICML 2026 · similarity 0.89)
• WebFactory: Automated Compression of Foundational Language Intelligence into Grounded Web Agents (ICLR 2026 · similarity 0.88)
• How to Train Your LLM Web Agent: A Statistical Diagnosis (NeurIPS 2025 · similarity 0.88)
• REAL: Benchmarking Autonomous Agents on Deterministic Simulations of Real Websites (NeurIPS 2025 · similarity 0.88)