Self-Evolving World Models for LLM Agent Planning

arXiv 2026 35.6 method

TLDR

WorldEvolver uses episodic and semantic memory to self-evolve world models for LLM agent planning, improving prediction accuracy and success rates.

Reasoning

The paper presents a novel framework that enhances world model reliability through test-time memory revision, with strong empirical results on two benchmarks. However, the abstract lacks details on limitations, scalability, or comparison to broader world model categories like generative or video world models.

Read-first score

Read-first score 35.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 33.

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,experiment

Topical relevance 29%

47.1

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=none

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

Frontier

Rank sensitivity

Stability: volatile; rank range: 153.

Keyword Scores

world model

world dynamics prediction

model-based reinforcement learning world model

interactive world model

world simulator

generative world model

video world model

Deep Analysis

Innovations

Self-evolving world model framework that revises deployment-time context while keeping agent and model parameters frozen
Integration of Episodic Memory (retrieval-based simulation from real transitions) and Semantic Memory (heuristic rules from prediction-observation mismatches)
Selective Foresight module that filters low-confidence predictions before integrating into agent reasoning

Methodology

WorldEvolver combines episodic memory for retrieval-based simulation, semantic memory for extracting heuristic rules from mismatches, and selective foresight to filter predictions. It is evaluated on ALFWorld and ScienceWorld using prediction accuracy (Word2World) and downstream agent success rate (AgentBoard), with frozen agent and model parameters.

Key Results

WorldEvolver achieves the highest prediction accuracy across three backbones and outperforms other world model baselines on downstream agent success rate.

Resources

Paper