Representation Learning for RL

RL: Reinforcement Learning

Representation learning for reinforcement learning studies how agents can turn observations and interaction histories into features that preserve what matters for predicting outcomes and choosing actions. Common approaches include predictive and contrastive latent models, successor and spectral representations, bisimulation metrics, and regularizers that encourage sparse, stable, or disentangled features. These methods support visual control, exploration, offline goal-conditioned navigation, world modeling, and transfer across tasks or environments. A central tension is that reusable features must discard irrelevant variation without losing details needed for control, and better-looking representations do not always improve downstream performance.

Latest papers

69 papers18 papers in the last four weeks, up 200% on the four weeks before. 0.1% of all new papers.

Next Embedding Prediction Makes World Models Stronger
Mar 3, 2026George Bredis, Nikita Balagansky, Daniil Gavrilov +1Model-Based RL Partially Observable RL
T-Tech.
Temporal Representations for Exploration: Learning Complex Exploratory Behavior without Extrinsic Rewards
Mar 2, 2026Faisal Mohamed, Catherine Ji, Benjamin Eysenbach +1RL Exploration Robotic RL
Mila-Quebec AI Institute · Université de Montréal · Princeton University
Sparse Masked Attention Policies for Reliable Generalization
Feb 23, 2026Caroline Horsch, Laurens Engwegen, Max Weltevrede +2Reinforcement Learning Generalization in RL
Department of Intelligent Systems, Delft University of Technology, Delft, Netherlands.
Stable Deep Reinforcement Learning via Isotropic Gaussian Representations
Feb 22, 2026Ali Saheb Pasand, Johan Obando-Ceron, Aaron Courville +2RL Stability Representation Learning for RL
Mila – Quebec AI Institute · McGill · Université de Montréal +1
Task-Aware Exploration via a Predictive Bisimulation Metric
Feb 21, 2026Dayang Liang, Ruihan Liu, Lipeng Wan +2RL Exploration Visual RL
Department of Automation, Xiamen University, Xiamen, China · Department of Computer Science, Beijing Normal-Hong Kong Baptist University, Zhuhai, China · School of Artificial Intelligence, Xian Jiaotong University, Xian, China +1
CDRL: A Reinforcement Learning Framework Inspired by Cerebellar Circuits and Dendritic Computational Strategies
Feb 17, 2026Sibo Zhang, Rui Jing, Liangfu Lv +2Reinforcement Learning Sample-Efficient RL
Academy of Medical Engineering and Translational Medicine, Tianjin University, Tianjin, China.
Can We Really Learn One Representation to Optimize All Rewards?
Feb 11, 2026Chongyi Zheng, Royina Karegoudra Jayanth, Benjamin EysenbachZero-Shot RL Representation Learning
Princeton University
BFM-Zero: A Promptable Behavioral Foundation Model for Humanoid Control Using Unsupervised Reinforcement Learning
Nov 6, 2025Yitang Li, Zhengyi Luo, Tonghe Zhang +10Humanoid Robot Control Behavioral Foundation Models
1Carnegie Mellon University · 2Meta
A Unified Framework for Zero-Shot Reinforcement Learning
Oct 23, 2025Jacopo Di Ventura, Jan Felix Kleuker, Aske Plaat +1Zero-Shot RL Formal Methods for RL
Leiden University

Mar 3, 2026J/K move · Enter open · S save

Next Embedding Prediction Makes World Models Stronger

George Bredis, Nikita Balagansky, Daniil Gavrilov, Ruslan Rakhimov

T-Tech.

Model-Based RL Partially Observable RL Representation Learning for RL Transformer-Based RL

Open paper AskPDF arXiv

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.