Masanari Oi

PhD student at Science Tokyo

profile.png

Okazaki Laboratory

Institute of Science Tokyo

I am a first-year PhD student in Computer Science at the Institute of Science Tokyo, and a member of the Okazaki Laboratory, where I work with Prof. Naoaki Okazaki. My research interests span natural language processing and computer vision. Recently, I have been particularly interested in embodied AI, especially spatial reasoning in vision-language models and vision-language navigation for robotics. I am also a member of Swallow-LLM, where we develop Japanese large language models.

I also work as a Research Engineer at CoeFont, where I develop text-to-speech systems, and at 3keigo, where I develop Japanese style-transfer systems for converting casual Japanese into keigo (敬語), a polite style of Japanese.

Recent News

Jul 14, 2026 A co-authored paper was accepted to COLM 2026 in San Francisco 🇺🇸.
Mar 31, 2026 Two first-authored papers were accepted to ICML 2026 in Seoul 🇰🇷.
Jan 31, 2026 A first-authored paper was accepted to LREC 2026 in Mallorca 🇪🇸.
Dec 31, 2025 A co-authored paper was accepted to ICLR 2026 in Rio de Janeiro 🇧🇷.
Oct 31, 2025 A co-authored paper was accepted to AAAI 2026 in Singapore 🇸🇬.

Selected Publications

Some of my earlier publications appear under the name Masanari Ohi, which was a previous romanization of my name. Both refer to myself, Masanari Oi.

2026

  1. hatch.jpg
    From Correspondence to Actions: Human-Like Multi-Image Spatial Reasoning in Multi-modal Large Language Models
    Masanari Oi, Koki Maeda, Ryuto Koike, and 3 more authors
    In ICML, 2026
    arXiv Code Website
    Proposed a training framework for vision‑language models to improve multi‑image spatial reasoning
  2. adpo.png
    Autoregressive Direct Preference Optimization
    Masanari Oi, Mahiro Ukai, Masahiro Kaneko, and 2 more authors
    In ICML, 2026
    arXiv Code Website
    Identified a theoretical issue in DPO and proposed a new method, ADPO.
  3. swallow_code.png
    Rewriting Pre-Training Data Boosts LLM Performance in Math and Code
    Kazuki Fujii, Yukito Tajima, Sakae Mizuki, and 14 more authors
    In ICLR 2026, 2026
    arXiv Dataset
    Proposed a rewriting‑based pipeline for synthesizing mathematics and coding data for LLMs.

2025

  1. hall_e.png
    HALL-E: Hierarchical Neural Codec Language Model for Minute-Long Zero-Shot Text-to-Speech Synthesis
    Yuto Nishimura, Takumi Hirose, Masanari Ohi, and 2 more authors
    In ICLR 2025, 2025
    arXiv
    Proposed an LLM‑based text‑to‑speech model for minute‑long zero-shot synthesis.

2024

  1. likelihood_bias.png
    Likelihood-based Mitigation of Evaluation Bias in Large Language Models
    Masanari Ohi, Masahiro Kaneko, Ryuto Koike, and 2 more authors
    In ACL Findings, 2024
    arXiv Code
    Detected and mitigated self‑preference bias in LLM‑as‑a‑judge evaluation.

Experiences

  • Research Assistant, Science Tokyo (Apr 2026 - Present)
  • AI Engineer, 3keigo.com (Apr 2023 - Present)
  • Research Engineer, CoeFont (Oct 2021 - Present)
  • Research Assistant, The National Institute of Advanced Industrial Science and Technology (AIST) (Aug 2024 - Mar 2026)

Education

  • Doctor of Engineering in Computer Science, Institute of Science Tokyo (Apr 2026 - Present)
  • Master of Engineering in Computer Science, Institute of Science Tokyo (Apr 2024 - Mar 2026)
    Formerly known as Tokyo Institute of Technology.
  • Bachelor of Engineering in Computer Science, Tokyo Institute of Technology (Apr 2020 - Mar 2024)

Honors

  • Best Paper Award, The 32nd Annual Meeting of the Association for Natural Language Processing (Mar 2026; Top 3 / 789 = 0.4%)
  • Best Paper Award, Journal of Natural Language Processing (Mar 2026; Top 2 / 33 = 6.1%)
  • Annual Conference Award, The 39th Annual Conference of the Japanese Society for Artificial Intelligence (2025; Top 35 / 1178 = 3.0%)
  • Committee Special Award, The 31st Annual Meeting of the Association for Natural Language Processing (Mar 2025; Top 63 / 726 = 8.7%)
  • Language Resource Award, The 31st Annual Meeting of the Association for Natural Language Processing (Top 1 / 42 = 2.4%)
  • Best Research Award, The 261st NL Research Presentation (Sep 2024; Top 1 / 15 = 6.7%)
  • Young Researcher Award, The 30th Annual Meeting of the Association for Natural Language Processing (Mar 2024; Top 18 / 427 = 4.2%)
  • Outstanding Paper Award, The 30th Annual Meeting of the Association for Natural Language Processing (Mar 2024; Top 13 / 599 = 2.1%)
  • Outstanding Paper Award, The 30th Annual Meeting of the Association for Natural Language Processing (Mar 2024; Top 13 / 599 = 2.1%)

Grants

  • Science Tokyo BOOST, Institute of Science Tokyo (Apr 2026 - Mar 2029)
    Scholarship: 3,600,000 JPY per year; research funds: 300,000 JPY per year.
  • Takenaka Scholarship, Takenaka Scholarship Foundation (Apr 2021 - Mar 2029)
    Scholarship: 1,200,000 JPY per year.
  • Full Repayment Exemption for Outstanding Academic Achievement, Japan Student Services Organization (JASSO) (Oct 2024 - Mar 2026)
    Full exemption from repayment of a JASSO student loan.

Academic Service

  • Reviewer, IEEE/ACM Transactions on Audio, Speech, and Language Processing (TASLP) (2025)
  • Emergency Reviewer, ACL Rolling Review (ARR), October cycle (2025)