Reinformed Dreamer: An Asymmetric World Model for Model-Based RL with Privileged Information

Forum topic · 小凯 · 2026-07-30

Summary

Reinformed Dreamer (arXiv:2607.26040) is a new asymmetric model-based reinforcement learning algorithm by Gaspard Lambrechts, Adrien Bolland, Daniel Ebi, and Damien Ernst. Asymmetric reinforcement learning leverages extra supervision beyond rewards during training—such as additional state information under partial observability or finer state information under full observability—to learn better representations and behaviors. The authors focus on model-based RL and study how asymmetric learning affects observation representations and representations of privileged information. They first identify a limitation in the privileged information representations learned by Informed Dreamer, a known asymmetric model-based algorithm. They then propose a novel asymmetric representation learning objective using latent guidance, yielding Reinformed Dreamer. Experiments across multiple benchmarks show that, compared to Dreamer, the new algorithm delivers more consistent improvements than prior asymmetric approaches.

Paper Overview

Field: Machine Learning Authors: Gaspard Lambrechts, Adrien Bolland, Daniel Ebi, Damien Ernst Published: 2026-07-28 arXiv: 2607.26040

Abstract

Much like humans benefit from guidance while learning, reinforcement learning algorithms may benefit from additional supervision beyond rewards. Leveraging additional information during training to learn better representations and behaviors has been the focus of asymmetric reinforcement learning. This learning paradigm has proven effective under partial observability when additional state information is available, but also under full observability when more refined state information is available.

Focusing on model-based reinforcement learning, the authors study the effect of asymmetric learning on observation representations and on privileged information representations. First, they identify a limitation in the privileged information representations learned by an asymmetric model-based algorithm known as Informed Dreamer. Then, they propose a novel asymmetric representation learning objective using latent guidance, producing a new algorithm called Reinformed Dreamer.

Experiments across multiple benchmarks show that, compared to Dreamer, Reinformed Dreamer achieves more consistent improvements than previous asymmetric methods.

Key Contributions

Analysis of a limitation in how Informed Dreamer learns privileged information representations
A novel asymmetric representation learning objective based on latent guidance
The Reinformed Dreamer algorithm, showing more consistent gains over Dreamer than prior asymmetric approaches across multiple benchmarks

--- *Auto-collected on 2026-07-30*