Agent2World: A Unified LLM-based Multi-Agent Framework for Symbolic World-Model Generation

ICLR 2026 Conference Submission17151 Authors

Keywords: Large Languge Model; LLM-based Agents; World Model

Abstract: Symbolic world models, which formally represent environment dynamics and constraints, are essential for model-based planning. While leveraging large language models (LLMs) to automatically generate these models from natural language has shown promise, existing approaches predominantly rely on scripted workflows that follow predetermined execution paths regardless of intermediate outcomes, often leading to inefficient computations and suboptimal solutions. In this paper, we propose Agent2World, a novel paradigm that employs autonomous tool-augmented LLM-based agents to generate symbolic world models adaptively. We further introduce Agent2World$_{\\text{Multi}}$, a unified multi-agent framework with specialized agents: (i) a Deep Researcher agent performs knowledge synthesis by web searching to address specification gaps; (ii) a Model Developer implements executable world models; and (iii) a specialized Testing Team conducts evaluation-driven refinement via systematic unit testing and simulation-based validation. Agent2World demonstrates superior performance across three benchmarks spanning both Planning Domain Definition Language(PDDL) and executable code representations, achieving consistent state-of-the-art results through a single unified framework. By enabling proactive, knowledge-grounded world-model generation, this work opens new possibilities for AI systems that can reliably understand and formalize complex environments.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 17151