Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
A machine learning approach that beats Rubik's cubes
Authors: Alexander Chervov, Kirill Khoruzhii, Nikita Bukhal, Jalal Naghiyev, Vladislav Zamkovoy, Ivan Koltsov, Lyudmila Cheldieva, Arsenii Sychev, Arsenii Lenin, Mark Obozov, Egor Urvanov, Alexey Romanov
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The paper reports on purely empirical research. Our results demonstrate that the proposed method is effective for multiple groups and outperforms all previously published machine learning solutions for solving Rubik s cubes up to 5x5x5. |
| Researcher Affiliation | Academia | Alexander Chervov Institut Curie, Universite PSL Paris, F-75005, France Kirill Khoruzhii Technical University of Munich Garching, 85748, Germany Nikita Bukhal Novosibirsk State University Novosibirsk, 630090, Russia Jalal Naghiyev Technical University of Munich Garching, 85748, Germany Vladislav Zamkovoy Independent Researcher Moscow, 119454, Russia Ivan Koltsov RTU MIREA Moscow, 119454, Russia Lyudmila Cheldieva RTU MIREA Moscow, 119454, Russia Arsenii Sychev RTU MIREA Moscow, 119454, Russia Arsenii Lenin RTU MIREA Moscow, 119454, Russia Mark Obozov Innopolis University Innopolis, 420500, Russia Egor Urvanov RTU MIREA Moscow, 119454, Russia Alexey M. Romanov RTU MIREA Moscow, 119454, Russia EMAIL |
| Pseudocode | No | The key steps of the proposed method are illustrated in the figure 1a and described below |
| Open Source Code | Yes | The original source code is attached to the paper as Supplementary Material. The last version can be found on Git Hub: https://github.com/khoruzhii/cayleypy-cube. |
| Open Datasets | Yes | Deep Cube A dataset [3] was used for 3x3x3 Rubik s cube, and the 2023 Kaggle Santa Challenge [50] dataset was used for 4x4x4 puzzle. |
| Dataset Splits | Yes | Creating the training set via random walks. (Diffusion distance.) Generate N random walk trajectories starting from a selected node. ... This set will serve as the training data. ... A new dataset of 1M examples was generated before each training epoch. |
| Hardware Specification | Yes | All of the experiments reported in the paper were performed on a dedicated server manufactured by Graviton. This server is equipped with two Intel Xeon Gold 6442Y 24-core processors running at 2600 Mhz, 256 GB DDR5 RAM, 512 GB SSD, and two GPUs NVIDIA H100 80GB. |
| Software Dependencies | Yes | The server was running Ubuntu Linux 24.04.2 LTS, CUDA 12.4. |
| Experiment Setup | Yes | The training procedure was performed using the Adam optimizer with a fixed learning rate of 0.001 and mean squared error as the loss function. A new dataset of 1M examples was generated before each training epoch. All models were pre-trained and remained unchanged during puzzle-solving. Training was conducted using 32-bit floating point precision, while inference used 16-bit floating point numbers to enhance computational efficiency. |