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

LS-Imagine: Open-World Reinforcement Learning over Long Short-Term Imagination

ICLR 25 Oral 2025 69.6 method

TLDR

LS-Imagine extends imagination horizon in open-world RL using a long short-term world model with goal-conditioned jumpy transitions, improving exploration and long-term feedback.

Reasoning

The paper introduces a novel method to extend imagination horizon via goal-conditioned jumpy transitions and affordance maps, showing strong empirical results on MineDojo. However, the approach relies on single-image zooming, which may limit generalization, and the abstract does not discuss potential limitations or failure cases.

Read-first score

Read-first score 69.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 49.

Recency 8%
86.7

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

Reproducibility 25%
73

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

Topical relevance 42%
70

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

Methodology quality 25%
60

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

Field roles

FrontierReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 74.

Keyword Scores

world model
10
model-based reinforcement learning world model
10
interactive world model
9
world dynamics prediction
7
generative world model
6
world simulator
5
video world model
2

Deep Analysis

Innovations

  • Extending imagination horizon within a limited number of state transition steps via a long short-term world model
  • Simulating goal-conditioned jumpy state transitions to enable long-horizon exploration
  • Computing affordance maps by zooming in on specific areas within single images to integrate direct long-term values into behavior learning

Methodology

LS-Imagine builds a long short-term world model that simulates goal-conditioned jumpy state transitions and computes affordance maps by zooming in on specific areas within single images. This allows the agent to integrate direct long-term values into behavior learning, extending the imagination horizon without requiring many transition steps. The method is evaluated in the MineDojo open-world environment against state-of-the-art techniques.

Key Results

LS-Imagine demonstrates significant improvements over state-of-the-art techniques in the MineDojo environment, indicating better exploration efficiency and long-horizon decision-making.

Limitations

  • Evaluation is limited to the MineDojo environment; generalization to other open-world domains is not demonstrated.
  • The approach relies on goal-conditioned jumpy state transitions, which may be non-trivial to define for tasks without clear subgoals.
  • Computing affordance maps by zooming in on specific areas may introduce computational overhead and may not scale well to very high-resolution images.

Tags