Generating Symbolic World Models via Test-time Scaling of Large Language Models

arXiv 25.2 2025 42 method

TLDR

LLMs generate symbolic world models (PDDL domains) via test-time scaling, outperforming o1-mini without additional training.

Reasoning

The paper introduces a novel method using Best-of-N sampling and verbalized ML to generate PDDL domains, achieving over 50% success rate on two tasks. Strengths include no need for extra training and clear outperformance of a strong baseline. Weaknesses are limited task scope and lack of real-world evaluation.

Read-first score

Read-first score 42, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 29.

Recency 8%

86.7

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

Topical relevance 42%

41.4

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%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=none

Reproducibility 25%

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: 361.

Keyword Scores

world model

world dynamics prediction

generative world model

model-based reinforcement learning world model

interactive world model

world simulator

video world model

Deep Analysis

Innovations

Scaling test-time computation of LLMs to enhance PDDL reasoning capabilities without additional training
Combining Best-of-N sampling with verbalized machine learning refinement for generating high-quality PDDL domains

Methodology

The method scales up test-time computation of LLMs to improve PDDL domain generation. It first uses Best-of-N sampling to produce an initial solution, then refines it in a fine-grained manner using verbalized machine learning. This approach does not require additional training and leverages PDDL as a formal planning abstraction.

Key Results

The method outperforms o1-mini by a considerable margin, achieving over 50% success rate on two tasks (generating PDDL domains from natural language description or PDDL problems). It also outperforms current state-of-the-art methods on almost all competition-level planning tasks.

Resources

Paper