StressDream: Steering Video World Models for Robust Policy Evaluation and Improvement

arXiv 2026 63.8 method, application

TLDR

StressDream steers diffusion-based video world models toward high-impact plausible outcomes by optimizing initial noise with semantic and plausibility objectives for robust policy evaluation.

Reasoning

The paper presents a novel method for steering video world models to generate high-impact plausible futures, with strong empirical results on autonomous driving and robotic manipulation. However, the abstract lacks explicit discussion of limitations or comparison to baselines, and the reliance on diffusion models may limit generalizability.

Read-first score

Read-first score 63.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 57.

Recency 6%

100

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

Citation impact 18%

92.4

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.9235999

Topical relevance 29%

81.4

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 18%

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

Reproducibility 18%

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

Citation velocity 12%

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

Field roles

FoundationFrontierBridgeMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 473.

Keyword Scores

world model

video world model

generative world model

world simulator

world dynamics prediction

model-based reinforcement learning world model

interactive world model

Deep Analysis

Innovations

Steering diffusion-based video world model imaginations toward high-impact plausible outcomes by optimizing initial noise at inference time
Two complementary objectives: a semantic objective using a Vision-Language Model for informative gradients, and a plausibility objective to prevent out-of-distribution noise
Enables robust policy evaluation and improvement by identifying actions whose plausible futures include undesirable outcomes (e.g., task failures) without requiring prohibitively many samples

Methodology

StressDream optimizes the initial noise of a diffusion-based video world model to steer generated imaginations toward text-specified high-impact outcomes (e.g., task failures). It uses a Vision-Language Model to provide semantic gradients by reasoning about the generated video, and a plausibility objective to keep the optimized noise within the in-distribution manifold. The method is evaluated on state-of-the-art video world models for autonomous driving and robotic manipulation tasks.

Key Results

StressDream effectively steers imaginations toward high-impact plausible outcomes specified by text at inference time, enabling robust policy evaluation and improvement by identifying actions whose plausible futures include undesirable outcomes.

Limitations

Optimization of high-dimensional noise remains challenging and may require careful balancing of the semantic and plausibility objectives
Relies on a pre-trained Vision-Language Model, which may introduce biases or fail to provide informative gradients in novel or complex scenes
Plausibility objective may not fully prevent out-of-distribution noise, potentially yielding implausible imaginations in edge cases
Evaluation is limited to autonomous driving and robotic manipulation; generalizability to other domains is not demonstrated

Resources

Paper