World Model Implanting for Test-time Adaptation of Embodied Agents

arXiv 25.9 2025 45.2 method, system

TLDR

WorMI combines LLMs with domain-specific world models via test-time composition for cross-domain adaptation in embodied AI.

Reasoning

The paper presents a novel framework (WorMI) that effectively integrates LLMs with retrieved world models for zero-shot and few-shot adaptation, validated on two benchmarks. However, it lacks real-world experiments and the abstract does not detail limitations or comparisons to non-LLM baselines.

Read-first score

Read-first score 45.2, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 26.

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,result

Topical relevance 42%

37.1

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 25%

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 357.

Keyword Scores

world model

interactive world model

model-based reinforcement learning world model

generative world model

world dynamics prediction

world simulator

video world model

Deep Analysis

Innovations

World model implanting framework (WorMI) that combines LLM reasoning with domain-specific world models via test-time composition
Prototype-based world model retrieval using efficient trajectory-based abstract representation matching
World-wise compound attention method that integrates knowledge from retrieved world models and aligns their intermediate representations with the reasoning model

Methodology

The WorMI framework enables test-time composition of an LLM-based reasoning model with independently learned, domain-specific world models. It uses a prototype-based retrieval approach that matches abstract trajectory representations to select relevant world models, and a world-wise compound attention mechanism to fuse domain-specific knowledge and align representations within the agent's policy.

Key Results

On VirtualHome and ALFWorld benchmarks, WorMI achieves superior zero-shot and few-shot performance compared to several LLM-based approaches across a range of unseen domains.

Resources

Paper