OPINE-World: Programmatic World Modeling with Ontology-error-Prioritized Interactive Exploration

arXiv 2026 42.2 method

TLDR

OPINE-World uses LLM agents to learn object-centric programmatic world models online via interactive exploration, evaluated on ARC-AGI-3 benchmark.

Reasoning

The paper presents a novel approach combining program synthesis with interactive exploration, addressing data efficiency and transfer. Strengths include a clear methodology and benchmark evaluation; weaknesses are limited scope (25 games) and potential scalability concerns.

Read-first score

Read-first score 42.2, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 41.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 18%

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

Topical relevance 29%

58.6

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

Reproducibility 18%

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

Citation impact 18%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 319.

Keyword Scores

world model

interactive world model

world dynamics prediction

model-based reinforcement learning world model

world simulator

generative world model

video world model

Deep Analysis

Innovations

Online learning of object-centric programmatic world models from interaction using LLM agents
Cooperative dual-agent loop: one acting in the environment, one synthesizing the model in code with replay verification and model-based planning
Exploration guided by a Bayesian measure of object-type adequacy called ontology error
Scaling program-synthesized world models to pixel-rendered environments without given object vocabulary

Methodology

OPINE-World is an LLM agent that learns an object-centric programmatic world model online by coupling two agents: one interacts with the environment, the other synthesizes a code model using replay verification and model-based planning. Exploration is steered by a Bayesian ontology error measure. Evaluation is on the ARC-AGI-3 benchmark where object vocabulary, goal, and action semantics are withheld.

Key Results

OPINE-World solves 20 of 25 ARC-AGI-3 games without per-game training and achieves an action-efficiency score of 78.4 against the human baseline.

Resources

Paper