One Model for All Tasks: Leveraging Efficient World Models in Multi-Task Planning

arXiv 25.9 2025 56.3 method

TLDR

Proposes ScaleZero, a multi-task world model using Mixture-of-Experts and dynamic parameter scaling to mitigate gradient conflicts and improve sample efficiency.

Reasoning

Strengths include systematic architectural investigation and theoretical/empirical validation of MoE for gradient conflict reduction, plus adaptive capacity allocation via DPS. Weaknesses: abstract is cut off, missing full results and limitations; reliance on simulated benchmarks (Atari, DMC, Jericho) rather than real-world tasks.

Read-first score

Read-first score 56.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 38.

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=benchmark,evaluation

Topical relevance 42%

54.3

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,github

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 242.

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

Mixture-of-Experts (MoE) architecture to mitigate gradient conflicts in multi-task world models
Online Dynamic Parameter Scaling (DPS) strategy using LoRA adapters for adaptive capacity allocation

Methodology

The paper extends the UniZero model with a Mixture-of-Experts (MoE) architecture to route task-specific representations to specialized sub-networks, theoretically and empirically alleviating gradient conflicts. Additionally, an online Dynamic Parameter Scaling (DPS) strategy progressively integrates LoRA adapters based on task-specific progress to dynamically allocate model capacity. The model is trained using online reinforcement learning on a diverse set of benchmarks (Atari, DMC, Jericho) and compared against specialized single-task agents.

Key Results

ScaleZero, using solely online reinforcement learning with one model, performs on par with specialized single-task agents. With the DPS strategy, it remains competitive while using only 71.5% of the environment interactions.

Resources

Paper