DraftFM: A FoundationModel for Day-Zero Drafting in Magic: The Gathering
DraftFM is a foundation model that enables day-zero drafting in Magic: The Gathering by scoring unseen cards using only their public attributes and draft context, achieving expert-level performance without relying on historical pick data or card identities.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the world of competitive card games, there is a unique challenge that arises the moment a new set of cards is revealed but before anyone has played with them. Imagine a group of players gathering to build decks from a fresh box of hundreds of new cards. They must decide, one by one, which cards to keep and which to pass to the next player, all while knowing nothing about how those cards will perform in actual games. This is the "day-zero" problem. For decades, computer programs designed to help players make these choices have been stuck waiting. They rely on learning from past games, but when a new set arrives, there are no past games to learn from. The data simply does not exist yet. This leaves players relying entirely on human intuition and expert opinions, which, while valuable, are subjective and vary from person to person. The question for researchers has long been whether a computer could learn the general rules of good card selection well enough to make smart guesses about a completely new set, without ever having seen a single pick from it.
A researcher named Brian Ward has tackled this problem with a new system called DraftFM. Instead of trying to memorize specific cards or waiting for data to accumulate, this system was trained on nearly 150 million choices made by human players across 29 different card sets over several years. The key innovation is how the computer sees the cards. Rather than treating each card as a unique name to be memorized, the system breaks every card down into a list of its physical and written characteristics: its color, its cost, its type, and the text of its rules. It also uses a fixed, pre-trained understanding of language to read the card's description. By learning from thousands of different sets, the system learned to recognize patterns in what makes a card valuable based on these features alone. It does not know the name of a card or which set it belongs to; it only knows the features it can see. This allows it to look at a brand-new card it has never encountered before and score it based on the same logic it used for cards it studied years ago.
To test if this approach actually works, the researcher held back three entire card sets from the training data. These sets were completely unknown to the model. When the system was asked to predict what human players would pick from these new sets, it performed with surprising accuracy. In the first pick of a draft, where a player has dozens of options and the chance of guessing correctly by random luck is only about 7 percent, the model agreed with human choices roughly 51 to 60 percent of the time. This is a significant leap forward, as previous computer models could not even attempt this task until after the sets had been played for weeks. The system also correctly identified that human players tend to be most consistent at the very beginning of a draft, when they have the most freedom to choose, and that their choices become more predictable as the pool of available cards shrinks.
The true test of this system came with the release of a new set based on The Hobbit. Before the set was even available for play on the digital platform where most people draft, the researcher used DraftFM to generate a complete ranking of every card in the set. This ranking was sealed and published with a digital timestamp roughly 36 hours before the set went live, ensuring that the predictions were made without any knowledge of how the cards would actually perform. The researcher then compared these computer-generated rankings against the pre-release reviews of six independent human experts who are well-known in the card community. The results showed that the computer's rankings agreed with the human experts about as much as the experts agreed with each other. While the computer did not outperform the humans, it managed to reach their level of consensus without having played a single game or seen a single draft log.
This work demonstrates that a computer can learn the underlying logic of a complex game well enough to make informed predictions about entirely new content. It proves that the ability to generalize from past experiences to future unknowns is possible even in a domain as intricate as card drafting. The system is not perfect; it is uniformly overconfident in its predictions, and it still struggles to match the nuance of human experts when the choices are difficult. However, it establishes a new baseline for what is possible on day zero. By treating cards as collections of features rather than memorized names, the model bridges the gap between learning from history and navigating the unknown. The researcher has committed to publishing a follow-up study once the actual draft data for the new set becomes available, which will reveal whether the model's predictions hold up against the reality of how the cards actually play. Until then, the work stands as a demonstration that machines can learn to play the game of prediction, even when the rules are being written for the first time.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.