EvoAgent: Agent Autonomous Evolution with Continual World Model for Long-Horizon Tasks

arXiv 25.2 2025 25.5 method

TLDR

EvoAgent uses a continual world model for autonomous long-horizon task completion in open worlds via self-planning, control, and reflection.

Reasoning

The paper addresses key challenges in long-horizon tasks with a novel curriculum self-evolving agent and continual world model, showing strengths in autonomous experience update and catastrophic forgetting mitigation. However, the abstract lacks details on world model architecture and comparisons, and experiments are limited to Minecraft.

Read-first score

Read-first score 25.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 15.

Recency 6%

86.7

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

Methodology quality 18%

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

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Topical relevance 29%

21.4

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

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: 58.

Keyword Scores

world model

model-based reinforcement learning world model

world dynamics prediction

world simulator

generative world model

interactive world model

video world model

Deep Analysis

Innovations

Curriculum self-evolving agent with a continual World Model for autonomous long-horizon task completion without human intervention
Experience-driven task planner using LLM and multimodal experiences to decompose long-horizon tasks into executable sub-tasks
World Model-guided action controller with self-verification mechanism for low-level action generation and multimodal experience update
Curriculum Learning-based reflector with a two-stage CL algorithm for task-adaptive selection of multimodal experiences to update the World Model
Planner-controller-reflector closed-loop dynamic enabling autonomous update of multimodal experiences and world knowledge

Methodology

EvolvingAgent comprises three modules: an experience-driven task planner that uses an LLM along with multimodal experiences to convert long-horizon tasks into sub-tasks; a World Model-guided action controller that generates low-level actions and incorporates a self-verification mechanism to update multimodal experiences; and a Curriculum Learning-based reflector that implements a two-stage CL algorithm to select multimodal experiences for task-adaptive World Model updates. These modules form a closed-loop dynamic for continual World Model evolution.

Key Results

On Minecraft, EvolvingAgent achieves a 111.74% improvement in average success rate and reduces ineffective actions by more than 6x compared to existing methods. It also generalizes to the Atari environment with human-level performance.

Resources

Paper