⌁ Awesome World Model Hub Papers · Datasets · Projects
中文 GitHub
← Back to papers

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

arXiv 25.7 2025 54.7 method

TLDR

Introduces SimuRA, an architecture using LLM-based world model for simulative reasoning, achieving up to 124% higher task completion in web-browser tasks.

Reasoning

Strengths: novel simulative reasoning architecture with LLM world model, significant improvements across diverse web tasks. Weaknesses: limited to web-browser environment, lacks comparison to other world model approaches or detailed limitations.

Read-first score

Read-first score 54.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 42.

Recency 8%
86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Topical relevance 42%
60

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

Methodology quality 25%
60

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

Reproducibility 25%
30

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 370.

Keyword Scores

world model
10
world simulator
8
world dynamics prediction
8
generative world model
6
model-based reinforcement learning world model
6
interactive world model
4
video world model
0

Deep Analysis

Innovations

  • Proposes simulative reasoning (System II) using an LLM-based world model with natural-language belief states for planning, replacing reactive decision-making (System I).
  • Introduces SiRA (Simulative Reasoning Architecture), a model-agnostic architecture that instantiates simulative reasoning for goal-oriented agents.
  • Demonstrates that counterfactual evaluation via world model simulation provides a generalizable planning mechanism across diverse task categories without task-specific tuning.

Methodology

The paper introduces SiRA, which uses an LLM-based world model with natural-language belief states to simulate future outcomes of candidate actions. The architecture is model-agnostic and evaluated across three task categories (constrained navigation, multi-hop information aggregation, general instruction following) in a web-browser environment. Baselines include a matched reactive agent and a representative open-web agent.

Key Results

Simulative reasoning achieves up to 124% higher task completion rates than the reactive baseline and increases constrained navigation success from 0% to 32.2% compared to the open-web agent. The persistent advantage across distinct task types suggests the benefit stems from generalizable counterfactual evaluation rather than task-specific tuning.

Resources

Paper

Tags