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

Aim Short to Reach Far: Your Frozen World Model Can Plan Better Than You Think

arXiv 2026 33.1 method

TLDR

Final-goal scoring in visual world model planning is suboptimal; aiming at intermediate observed targets improves control with frozen LeWM models.

Reasoning

The paper clearly identifies a limitation of final-goal scoring and introduces Anchored Planning, showing consistent improvements across multiple benchmark tasks. Strengths include a simple, well-motivated method and empirical evaluation, but limitations include reliance on recorded segments and no evidence from physical real-world systems.

Read-first score

Read-first score 33.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 35.

Recency 5%
100

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

Methodology quality 15%
60

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

Topical relevance 25%
50

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 15%
38

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

Graph prestige 15%
5.7

PageRank-style prestige over the local citation graph built from available OpenAlex references. local_pagerank=0.0016

Citation impact 15%
0

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 10%
0

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 253.

Keyword Scores

world model
9
world dynamics prediction
6
model-based reinforcement learning world model
6
world simulator
4
video world model
4
generative world model
3
interactive world model
3

Deep Analysis

Innovations

  • Shows that scoring predicted outcomes by distance to a final encoded goal can limit control even with exact dynamics and globally optimal short-horizon search, since reaching a goal may require actions that initially move away from it.
  • Demonstrates that intermediate targets substantially improve action synthesis and recorded-action ranking with frozen LeWM models across Cube, PushT, Reacher, and TwoRoom.
  • Introduces Anchored Planning, which retrieves a recorded segment whose start and end resemble current and goal observations, then aims at an observation shortly after its start using a frozen model.
  • Finds that planning toward observed targets without additional training outperforms the released LeWM planner on every task in long-range evaluation, while additional final-goal search falls short.
  • Identifies that lower successor-prediction error does not necessarily improve control, and that target distance ahead and shrinking retrieval span matter for success.

Methodology

The paper studies visual world model planners using frozen LeWM models on Cube, PushT, Reacher, and TwoRoom. It compares final-goal scoring against planning toward intermediate targets, including learned targets and targets drawn from observed experience. The proposed Anchored Planning retrieves a recorded segment matching current and goal observations, aims at an observation shortly after segment start, scores actions toward that target from the current state, and shrinks the retrieval span as execution advances.

Key Results

Intermediate targets substantially improve action synthesis and recorded-action ranking across Cube, PushT, Reacher, and TwoRoom, with gains from both learned and observed targets. Anchored Planning with observed targets outperforms the released LeWM planner on every task in long-range evaluation without additional training, while additional final-goal search does not match these gains.

Limitations

  • Additional final-goal search falls short of the same gains.
  • Lower successor-prediction error need not translate into better control.
  • Success depends on how far ahead the target is placed and on shrinking the retrieval span as execution advances.

Resources

Paper

Tags