日本語

Momose Oyama

I am a second-year Ph.D student at Shimodaira Lab, Kyoto University.
Research interest:
- Understanding How Embedding Geometry Encodes Meaning
- Interpretation of Neural Network Models
E-Mail: oyama.momose [atmark] sys.i.kyoto-u.ac.jp
Google Scholar

Publications

Momose Oyama, Ryo Kishino, Hiroaki Yamagiwa, Hidetoshi Shimodaira.
Likelihood Variance as Text Importance for Resampling Texts to Map Language Models
[arXiv] EMNLP 2025 Findings
Momose Oyama, Hiroaki Yamagiwa, Yusuke Takase, Hidetoshi Shimodaira.
Mapping 1,000+ Language Models via the Log-Likelihood Vector
[arXiv] [ACL Anthology] ACL 2025 🏆 Outstanding Paper Award
Hiroaki Yamagiwa, Momose Oyama, Hidetoshi Shimodaira.
Revisiting Cosine Similarity via Normalized ICA-transformed Embeddings
[arXiv] [ACL Anthology] COLING 2025
Momose Oyama, Hiroaki Yamagiwa, Hidetoshi Shimodaira.
Understanding Higher-Order Correlations Among Semantic Components in Embeddings
[arXiv] [ACL Anthology] EMNLP 2024
Hiroaki Yamagiwa*, Momose Oyama*, Hidetoshi Shimodaira.
Discovering Universal Geometry in Embeddings with ICA
[arXiv] [ACL Anthology] EMNLP 2023
Momose Oyama, Sho Yokoi, Hidetoshi Shimodaira.
Norm of Word Embedding Encodes Information Gain
[arXiv] [ACL Anthology] EMNLP 2023

Ryo Kishino, Yusuke Takase, Momose Oyama, Hiroaki Yamagiwa, Hidetoshi Shimodaira.
Revealing Language Model Trajectories via Kullback-Leibler Divergence
[arXiv]

2023/06 - Present Research Assistant, RIKEN Center for Advanced Intelligence Project
2022/04 - 2023/05 Research Assistant, Kyoto University
2021/07 - 2022/03 Research Assistant, RIKEN Center for Advanced Intelligence Project