Developer Releases Doomformer: Compiling Integer Neural Networks into DOOM Maps Achieves 97.16% MNIST Accuracy
Open-source project Doomformer runs neural networks inside DOOM maps using game mechanics, achieving 97.16% accuracy on MNIST digits with zero-tolerance verification against NumPy.
GitHub user mottopanikeiku has released Doomformer, an open-source project that compiles integer neural networks into map files for the classic first-person shooter DOOM. According to t3n.de and the GitHub repository, the experiment uses "Voodoo Dolls" (cloned player entities spawned when entering specific areas) as computational units, storing values in light levels and floor heights to perform neural network inference.
In tests on the MNIST handwritten digit dataset, Doomformer used a map network containing 1,355 Voodoo Dolls to identify 10,000 test samples, correctly recognizing 9,716 digits for an accuracy of 97.16%. The float model version achieved 97.58%. The project emphasizes strict verification: every value computed by the game engine is compared against integer reference results from NumPy with zero tolerance.
The core innovation lies in its compiler technique, which maps neural network layers to game map elements: input values correspond to crowds of dolls, weights to lines crossed by dolls, activation functions (ReLU) to staircase gates dolls must climb, and attention mechanisms to dolls squeezing under lowered ceilings. While demonstrating the feasibility of non-traditional computational substrates, the method is extremely slow—a small MLP runs about 23,000 times slower than NumPy—and currently supports only integer models without modern architecture components like LayerNorm.