irodkin/1dCA_r2s20T20

This dataset accompanies the paper Beyond Memorization: Extending Reasoning Depth with Recurrence, Memory and Test-Time Compute Scaling.

It provides a controlled cellular-automata (1dCA) framework designed to evaluate rule inference and multi-step reasoning depth in neural architectures, excluding the possibility of memorization by using disjoint training and test rules.

Resources

• Paper: arXiv:2508.16745
• Hugging Face Paper Page: HF Paper
• GitHub Repository: associative-recurrent-memory-transformer

Cite

@misc{rodkin2025memorizationextendingreasoningdepth,
      title={Beyond Memorization: Extending Reasoning Depth with Recurrence, Memory and Test-Time Compute Scaling}, 
      author={Ivan Rodkin and Daniil Orel and Konstantin Smirnov and Arman Bolatov and Bilal Elbouardi and Besher Hassan and Yuri Kuratov and Aydar Bulatov and Preslav Nakov and Timothy Baldwin and Artem Shelmanov and Mikhail Burtsev},
      year={2025},
      eprint={2508.16745},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2508.16745}, 
}