Saved in:
Bibliographic Details
Main Authors: Bredis, George, Balagansky, Nikita, Gavrilov, Daniil, Rakhimov, Ruslan
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2603.02765
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866917309121036288
author Bredis, George
Balagansky, Nikita
Gavrilov, Daniil
Rakhimov, Ruslan
author_facet Bredis, George
Balagansky, Nikita
Gavrilov, Daniil
Rakhimov, Ruslan
contents Capturing temporal dependencies is critical for model-based reinforcement learning (MBRL) in partially observable, high-dimensional domains. We introduce NE-Dreamer, a decoder-free MBRL agent that leverages a temporal transformer to predict next-step encoder embeddings from latent state sequences, directly optimizing temporal predictive alignment in representation space. This approach enables NE-Dreamer to learn coherent, predictive state representations without reconstruction losses or auxiliary supervision. On the DeepMind Control Suite, NE-Dreamer matches or exceeds the performance of DreamerV3 and leading decoder-free agents. On a challenging subset of DMLab tasks involving memory and spatial reasoning, NE-Dreamer achieves substantial gains. These results establish next-embedding prediction with temporal transformers as an effective, scalable framework for MBRL in complex, partially observable environments.
format Preprint
id arxiv_https___arxiv_org_abs_2603_02765
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Next Embedding Prediction Makes World Models Stronger
Bredis, George
Balagansky, Nikita
Gavrilov, Daniil
Rakhimov, Ruslan
Machine Learning
Artificial Intelligence
Capturing temporal dependencies is critical for model-based reinforcement learning (MBRL) in partially observable, high-dimensional domains. We introduce NE-Dreamer, a decoder-free MBRL agent that leverages a temporal transformer to predict next-step encoder embeddings from latent state sequences, directly optimizing temporal predictive alignment in representation space. This approach enables NE-Dreamer to learn coherent, predictive state representations without reconstruction losses or auxiliary supervision. On the DeepMind Control Suite, NE-Dreamer matches or exceeds the performance of DreamerV3 and leading decoder-free agents. On a challenging subset of DMLab tasks involving memory and spatial reasoning, NE-Dreamer achieves substantial gains. These results establish next-embedding prediction with temporal transformers as an effective, scalable framework for MBRL in complex, partially observable environments.
title Next Embedding Prediction Makes World Models Stronger
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2603.02765