Is Your LLM Secretly a World Model of the Internet? Model-Based Planning for Web Agents

arXiv 24.11 2024 68.2 method, application

TLDR

Proposes WebDreamer, a model-based planning framework using LLMs as world models for efficient web agent planning on real websites.

Reasoning

The paper introduces a novel model-based planning approach for web agents, demonstrating strong empirical results on both sandbox and real-world environments with significant efficiency gains. However, the reliance on LLMs as world models may introduce limitations in accuracy and generalizability not fully addressed.

Read-first score

Read-first score 68.2, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 44.

Recency 8%

75.1

Uses a gentle age decay so recent papers surface without erasing older foundations. 2024

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=True; dataset=False; markers=github

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=baseline,evaluation,result

Topical relevance 42%

62.9

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Field roles

Methodology anchorReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 142.

Keyword Scores

world model

world simulator

world dynamics prediction

generative world model

interactive world model

model-based reinforcement learning world model

video world model

Deep Analysis

Innovations

Proposing a model-based planning framework, WebDreamer, that employs LLMs as both world models and value functions.
Training specialized LLMs as world models with a scalable data synthesis pipeline.
Demonstrating that model-based planning is competitive with tree search while being 4-5 times more efficient in sandbox environments and effective on real-world websites.

Methodology

The paper proposes WebDreamer, a model-based planning framework for web agents that uses LLMs as world models to simulate action outcomes and as value functions. Specialized LLMs (Dreamer-7B) are trained using a scalable data synthesis pipeline. Evaluation is conducted on VisualWebArena (sandbox), Online-Mind2Web, and Mind2Web-Live (real-world) against reactive baselines and tree search, measuring performance and efficiency.

Key Results

WebDreamer achieves substantial performance improvements over reactive baselines, and is competitive with tree search while being 4-5 times more efficient in sandbox environments. It also works effectively on real-world websites, with the trained Dreamer-7B performing comparably to GPT-4o.

Resources

Paper Code