One Life to Learn: Inferring Symbolic World Models for Stochastic Environments from Unguided Exploration

arXiv 25.10 2025 60.1 method

TLDR

Introduces OneLife, a framework for learning symbolic world models in stochastic environments with limited interaction, using conditionally-activated programmatic laws.

Reasoning

The paper addresses a challenging setting (stochastic, one-life exploration) and proposes a novel probabilistic programming approach. Strengths include a new evaluation protocol and handling of sparse rule activation; weaknesses include reliance on a simulated environment (Crafter-OO) and lack of real-world validation.

Read-first score

Read-first score 60.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 35.

Methodology quality 25%

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

Recency 8%

86.7

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

Topical relevance 42%

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 25%

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

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 316.

Keyword Scores

world model

world dynamics prediction

generative world model

model-based reinforcement learning world model

world simulator

interactive world model

video world model

Deep Analysis

Innovations

OneLife framework that models world dynamics through conditionally-activated programmatic laws within a probabilistic programming framework
Dynamic computation graph that routes inference and optimization only through relevant laws, avoiding scaling challenges
New evaluation protocol measuring state ranking (distinguishing plausible from implausible future states) and state fidelity (generating realistic future states)
Crafter-OO, a reimplementation of the Crafter environment with a structured, object-oriented symbolic state and pure transition function

Methodology

OneLife uses conditionally-activated programmatic laws with precondition-effect structures in a probabilistic programming framework. It builds a dynamic computation graph that routes inference and optimization only through laws relevant to the current state, enabling learning of stochastic dynamics with sparse activation. The framework is evaluated on Crafter-OO using minimal unguided interaction data, with a strong baseline (likely neural network) and metrics of state ranking and state fidelity.

Key Results

OneLife outperforms a strong baseline on 16 out of 23 scenarios in state ranking and state fidelity. Simulated rollouts using the learned model successfully identify superior strategies, demonstrating planning ability.

Limitations

Relies on a structured, object-oriented symbolic state representation, limiting applicability to environments without such structure
Evaluation is conducted only on Crafter-OO, so generalization to other complex stochastic environments is not demonstrated
The 'one life' constraint may result in incomplete coverage of state space and learning only key dynamics, potentially missing rare or complex transitions
Scalability to larger state spaces or more complex hierarchical dynamics is not addressed

Resources

Paper Project