Dreaming Smoothly and Sample Efficiently with Gradient Penalized Latent Dynamics

arXiv 2026 60.2 method

TLDR

GPLD regularizes DreamerV3's latent dynamics with a Jacobian penalty to improve sample efficiency in continuous control tasks.

Reasoning

The paper presents a simple yet effective regularization technique that improves sample efficiency, especially on complex locomotion tasks. However, the evaluation is limited to simulated proprioceptive tasks, and the method's generality to other domains (e.g., visual or discrete) is not demonstrated.

Read-first score

Read-first score 60.2, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 38.

Recency 6%

100

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

Reproducibility 18%

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

Citation impact 18%

76.2

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.76219694

Methodology quality 18%

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

Topical relevance 29%

54.3

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

Citation velocity 12%

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

Field roles

FoundationFrontierBridgeReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 471.

Keyword Scores

model-based reinforcement learning world model

world model

world dynamics prediction

generative world model

world simulator

interactive world model

video world model

Deep Analysis

Innovations

Gradient-penalized latent dynamics regularizer (GPLD) for DreamerV3
Row-wise Jacobian penalty on the posterior latent distribution to encourage locally smooth transition learning
Interpretation of the penalty as the continuous-latent analog of finite-difference smoothing in discrete embedded-state MDPs
Efficient estimation of the penalty using Hutchinson-style stochastic probes

Methodology

The authors propose GPLD, a regularizer that adds a row-wise Jacobian penalty to the posterior latent distribution of DreamerV3's latent dynamics. The penalty is estimated efficiently via Hutchinson stochastic probes. The method is evaluated on DeepMind Control proprioceptive tasks and quadruped locomotion tasks, comparing sample efficiency and learning consistency against the DreamerV3 baseline.

Key Results

GPLD improves aggregate sample efficiency across DeepMind Control tasks, with particularly strong gains on higher-complexity locomotion environments. On quadruped tasks, GPLD reaches high-return behavior earlier and exhibits more consistent late-stage learning over longer horizons.

Limitations

The method is demonstrated only on smooth continuous control environments; its effectiveness on discrete or non-smooth environments is not evaluated.
The computational overhead of the Hutchinson-style stochastic probes is not quantified in the abstract.

Resources

Paper Code