⌁ Awesome World Model Hub Papers · Datasets · Projects
中文 GitHub
← Back to papers

Imagine-then-Plan: Agent Learning from Adaptive Lookahead with World Models

arXiv 26.1 2026 59.5 method

TLDR

Proposes Imagine-then-Plan, a framework with adaptive lookahead using world models for agent planning, outperforming baselines on benchmarks.

Reasoning

Strengths include a novel adaptive lookahead mechanism and unified framework with both training-free and RL variants, supported by extensive experiments. Weaknesses are lack of specific benchmark details and limitations discussion in the abstract.

Read-first score

Read-first score 59.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 33.

Recency 8%
100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 25%
80

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

Topical relevance 42%
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 25%
46

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=code,github

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 446.

Keyword Scores

world model
10
world dynamics prediction
7
model-based reinforcement learning world model
6
generative world model
4
world simulator
3
interactive world model
2
video world model
1

Deep Analysis

Innovations

  • Unified framework for agent learning via lookahead imagination (ITP) that generates multi-step imagined trajectories by interacting a policy model with a learned world model.
  • Adaptive lookahead mechanism that dynamically adjusts the imagination horizon by trading off the ultimate goal and task progress.
  • Formulation of a partially observable and imaginable Markov decision process (POIMDP) that fuses imagined future signals with current observations to guide policy learning.
  • Two instantiations of ITP: a training-free variant and a reinforcement-trained variant.

Methodology

ITP uses a learned world model to enable an agent's policy model to generate multi-step imagined trajectories. An adaptive lookahead mechanism adjusts the imagination horizon based on task progress and goal trade-offs. The imagined trajectories provide future signals (e.g., progress, conflicts) that are fused with current observations to form a partially observable and imaginable MDP, which then guides policy learning. The framework is instantiated in both training-free and reinforcement-trained variants.

Key Results

ITP significantly outperforms competitive baselines across representative agent benchmarks. Further analyses confirm that the adaptive lookahead mechanism substantially enhances the agent's reasoning capability.

Resources

Paper

Tags