Hello! I am a final-year PhD student (est. March 2026) at the Institute of Science Tokyo, advised by Prof. Naoaki Okazaki. During my PhD, I joined Prof. Chris Callison-Burch lab at Penn NLP as a Visiting Researcher in 2024. I also work closely with Prof. Preslav Nakov at MBZUAI NLP. Outside of academia, I am involved as a research advisor for a startup on multi-lingual text generation.
|
Email: ryuto.koike [at] nlp.c.titech.ac.jp 
|
|
|
I am passionate about making AI systems actually work in the real world, with my primary research
focus on Trustworthy NLP.
My work includes model robustness, the creation of interpretable NLP systems, ensuring
the safe
application of LLMs, and rigorous evaluation grounded in practical scenarios.
Currently, I am exploring the automated detection of AI‑generated content, specifically on
building deployable detectors in practical scenarios with minimum harm. |
|
|
|
|
Robustness
Ryuto Koike, Masahiro Kaneko, Naoaki Okazaki AAAI 2024
TL;DR - We propose OUTFOX, a novel framework that improves the robustness of LLM text detectors by
allowing both the detector and the attacker to adversarially learn from each other's output through
in-context learning. In this framework, the attacker uses the detector's prediction labels as examples
for in-context learning and adversarially generates essays that are harder to detect, while the
detector uses the adversarially generated essays as examples for in-context learning to learn to
detect essays from a strong attacker.
🏆 Double
Sponsorship Awards (1/140 ≈ 0.7%) in YANS
📸 Featured in Nikkei, NAACL Tutorial, Originality.ai Blog 📈 +40 stars in GitHub and +90 citations in Google Scholar |
|
Robustness Evaluation
Ryuto Koike, Masahiro Kaneko, Naoaki Okazaki EMNLP 2024 (Findings)
TL;DR - We reveal that current detectors are brittle to instruction variation in text
generation.
Specifically, even task-oriented constraints -- constraints that would naturally be included in
an
instruction and are not related to detection-evasion -- cause existing powerful detectors to have a
large variance in detection performance.
We raise awareness of the need to ensure prompt diversity when creating a detection benchmark and
open-source our constrained dataset.
|
|
Interpretability Reliability
Ryuto Koike, Masahiro Kaneko, Ayana Niwa, Preslav Nakov, Naoaki Okazaki NeurIPS (In submission)
TL;DR - We propose ExaGPT, an interpretable AI text detector that identifies a text by checking
whether it shares more similar spans with human-written vs. machine-generated texts from a datastore
and presents those spans as evidence for users to assess how reliably correct the decision is.
|
|
Reliability Evaluation
Masanari Ohi†, Masahiro Kaneko, Ryuto Koike, Mengsay Loem, Naoaki Okazaki ACL 2024 (Findings)
TL;DR - We present a self-preference bias in LLM-as-a-judge i.e., LLMs overrate texts with higher
likelihoods while underrating those
with lower likelihoods. We also propose a simple but effective mitigation method via in-context
learning.
🏆 Outstanding
Young Researcher’s Paper (18/427 ≈ 4.2%) in ANLP
|
|
|
|
|
University of Pennsylvania, Philadelphia, PA, USA Visiting Researcher (2024.10 - 2025.04) Advisor: Prof. Chris Callison-Burch |
|
Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), Abu Dhabi, UAE Research Collaborate (2024.04 - Present) Advisor: Prof. Preslav Nakov |
|
Exawizards, Inc., Tokyo, Japan Machine Learning Engineer Intern (2022.02 - 2022.03) |
|
CyberAgent, Inc., Tokyo, Japan Research Intern (2021.09 - 2022.01) Software Engineer Intern (2021.07 - 2021.08) |
|
|
|
|
Institute of Science Tokyo (Formerly, Tokyo Tech), Tokyo, Japan Ph.D. in Computer Science (2023.04 - est. 2026.04) |
|
Keio University, Tokyo, Japan M.S. in Information and Computer Science (2021.04 - 2023.03) B.S. in Information and Computer Science (2017.04 - 2021.03) |
|
|
|
|
©︎ Ryuto Koike / Design: Jon Barron. |