Original title: NashDreamer: Model-Based Reinforcement Learning for Zero-Sum Imperfect-Information Games
Tomáš Holeček, Viliam Lisý · September 2026
Games where players keep secrets from each other, like poker, are tough for artificial intelligence. When opponents change how they play, the game constantly shifts. Learning a model of the game separately for each player hits major mathematical roadblocks. To fix this, researchers created NashDreamer for two-player games where one player's gain is the other's loss. NashDreamer uses one shared, centralized model of the world. This model separates the true game rules from how player choices change what each person gets to see. NashDreamer can use standard learning algorithms and can guarantee finding balanced strategies under an idealized model. Across four test games, NashDreamer learned much faster than methods that do not model the world, needing far fewer practice trials early on.
Imagine playing a game of charades where each team can only see their own clues and guesses, but not the other team's. Trying to learn a winning strategy is tough because you don't know what the other team is thinking or doing. In this paper, the system acts like a master referee who sees all the clues and actions across both teams, using that full view to help each team learn better strategies without revealing secrets directly.
This helps AI researchers and game developers train models for complex games faster and more effectively, potentially improving AI performance in strategic applications like poker or defense simulations.
The abstract notes vulnerability to posterior collapse in stochastic environments, which is left as an open challenge. It has only been tested on four benchmark games and is designed specifically for two-player zero-sum games, so generalizability is not established.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
NashDreamer: Model-Based Reinforcement Learning for Zero-Sum Imperfect-Information Games — Tomáš Holeček, Viliam Lisý
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.