# SimuRA: Towards General Goal-Oriented Agent via Simulative Reasoning Architecture with LLM-Based World Model

**Authors:** Mingkai Deng, Jinyu Hou, Yilin Shen, Hongxia Jin, Graham Neubig, Zhiting Hu, Eric Xing  
**arXiv:** [2507.23773](https://arxiv.org/abs/2507.23773)  
**Date:** 2025-07-31  
**Summary page:** https://www.summarizepaper.com/en/paper/2507.23773v1/

## Key points

- AI agents built on large language models (LLMs) have shown great promise
- Current approaches often focus on a one-task-one-agent model lacking scalability and generality
- Humans are general problem-solvers who can reason and plan across diverse environments by simulating outcomes
- Introduction of SimuRA (Simulative Reasoning Architecture), a goal-oriented framework for generalized agentic reasoning
- SimuRA leverages a world model for planning through simulation, overcoming constraints of autoregressive LLMs
- Experiments show success rate improvement in flight searches using SimuRA's world-model-based planning
- SimuRA architecture includes policy module, world model, and critic module for action selection based on goals and outcomes evaluation
- Natural language used as a compact representation for simulation in SimuRA ensures robustness and adaptability across tasks
- SimuRA available as an open-source library through LLM Reasoners with REASONERAGENT-WEB serving as research preview
- Ongoing efforts to expand the system to tackle broader challenges and showcase versatility across different task domains

## Summary

AI agents built on large language models (LLMs) have shown great promise, but current approaches often focus on a one-task-one-agent model that lacks scalability and generality. These agents also face limitations inherent in autoregressive reasoning. In contrast, humans are general problem-solvers who can reason and plan across diverse environments by simulating outcomes and planning accordingly. To address these challenges, we introduce SimuRA (Simulative Reasoning Architecture), a goal-oriented framework for generalized agentic reasoning. By leveraging a world model for planning through simulation, SimuRA overcomes the constraints of autoregressive LLMs. This world model is implemented using LLMs, allowing for flexible planning in various environments using the rich latent space of natural language. Experiments conducted on challenging web browsing tasks demonstrate the effectiveness of SimuRA. The success rate of flight searches improved from 0% to 32.2%, with world-model-based planning consistently outperforming autoregressive planning by up to 124%. This highlights the advantage of simulation-based reasoning as a paradigm for AI agents. The architecture of SimuRA involves a policy module that proposes potential actions based on goals, a world model that simulates outcomes, and a critic module that evaluates these outcomes to select the best action. By utilizing natural language as a compact representation for simulation, SimuRA ensures robustness and adaptability across tasks. We have made SimuRA available as an open-source library through LLM Reasoners, with the web agent REASONERAGENT-WEB serving as a research preview. Ongoing efforts are focused on expanding the system to tackle broader challenges and showcase its versatility across different task domains. Overall, our results demonstrate that SimuRA offers significant improvements over baseline approaches in complex website navigation tasks. The architecture's ability to reason through simulation shows promise for developing more general and powerful AI agents capable of superintelligent performance across diverse environments.