CoEx -- Co-evolving World-model and Exploration

arXiv 25.7 2025 48.3 method

TLDR

CoEx introduces a hierarchical agent architecture where LLM planning co-evolves with a dynamically updated neurosymbolic world model, outperforming baselines on ALFWorld, PDDL, and Jericho.

Reasoning

The paper presents a novel co-evolution mechanism for world model updates, addressing a key limitation of static LLM world models. Strengths include hierarchical state abstraction and empirical validation across diverse environments. Weaknesses include reliance on LLM reasoning and limited discussion of scalability or failure cases.

Read-first score

Read-first score 48.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 32.

Recency 8%

86.7

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

Methodology quality 25%

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

Topical relevance 42%

45.7

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=code

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 392.

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

Co-evolving world-model and exploration via hierarchical agent architecture
Hierarchical state abstraction allowing LLM planning to co-evolve with dynamically updated world model
Neurosymbolic belief state combining textual inferences and code-based symbolic memory
Learning mechanism that continuously incorporates subgoal experiences into persistent world model

Methodology

CoEx uses a hierarchical agent architecture where LLM reasoning orchestrates dynamic plans consisting of subgoals. A learning mechanism continuously incorporates subgoal experiences into a persistent world model in the form of a neurosymbolic belief state, comprising textual inferences and code-based symbolic memory. The agent is evaluated across diverse environments including ALFWorld, PDDL, and Jericho.

Key Results

CoEx outperforms existing agent paradigms in planning and exploration across the evaluated environments (ALFWorld, PDDL, Jericho).

Resources

Paper