Paper Accepted at NeurIPS 2025, One of the World’s Premier Conferences in AI

A paper by Tsubasa Masumura and Associate Professor Masato Taki (Visiting Scientist at iTHEMS), both from Rikkyo University, has been accepted for presentation at NeurIPS 2025, an international conference in the field of machine learning.
In this study, the authors proposed a new Transformer architecture that incorporates the principles of associative memory, successfully mitigating the issue known as “rank collapse” both theoretically and experimentally.
The proposed method improves model performance without introducing additional parameters, offering potential contributions to the further advancement of large language models and generative AI.
NeurIPS 2025 will be held in December 2025 in San Diego, USA.

For more details, please refer to the Rikkyo University press release via the related link below.