San Jose, CA Cambridge, MA Chiba, Japan
In the past, I had the pleasure of being advised by Hidenori Tanaka, Ekdeep Singh Lubana, and Hanspeter Pfister at Harvard, as well as Tobias Höllerer at UCSB. Through my undergraduate and ongoing graduate studies, the Ezoe Memorial Recruit Foundation Scholarship has graciously supported my academic pursuits.
My research interest is to understand the surprising quirks of deep learning. Why does training give birth to well-organized representations for certain concepts and tasks, but not others? What mechanistic motifs emerge across different models, and why? What properties of the underlying learning algorithm and data distribution lead to these phenomena, and how can we leverage this understanding to build safer and more capable systems? I want to answer these fundamental questions by building a scientific theory of artificial intelligence. Incidentally, I strongly support interdisciplinary collaboration, open access, and open source.
Aside from academics, I'm an avid long-distance runner (mainly half and full marathons). I also love F1, public transit, anime/vtubers, music production, local eats, and Rocket League. Feel free to reach out via email or Discord @kento24!
Preprint 2026, as first author.
We show that generative sequence models can produce individually plausible physical trajectories while shifting aggregate quantities such as distance or energy, and use a data deviation kernel to predict and reduce this drift.
ICML 2026 SPIGM Workshop, as co-author.
We train biological sequence models with a curriculum that gradually expands from nearby evolutionary neighbors to more distant homologs, improving protein variant-effect prediction and RNA sequence generation.
Sci4DL 2026 Workshop, as first author.
We show that generated samples can fit the observed data distribution while violating the distribution of hidden controls such as speed, energy, or speaking rate.
ICML 2025, as first author.
Poster/OpenReview/arXiv/Code/Tweet Thread
We show that knowledge editing distorts entity representations beyond targeted facts, fracturing geometries that support factual recall and reasoning.
ICLR 2025, as co-author.
We show that a concept representation can mediate in-context learning; when a graph is specified by enough examples, model representations shift toward that graph rather than only pretrained semantic associations.
NeurIPS 2024 NeurReps Workshop, as co-author.
We find internal task representations during in-context learning on synthetic sequences from geometric graphs, including cases where model behavior follows context-specified structure over semantic priors.
ICML 2024, as co-author.
We use a synthetic graph-navigation task to measure stepwise inference, covering the reasoning gap, sampling-temperature tradeoff, simplicity bias, compositional generalization, and in-context primacy.
NeurIPS 2023 R0-FoMo Workshop, as co-author.
In directed acyclic graph navigation, we find better route generation when step-by-step paths expose hierarchical subpaths seen during pretraining; generated routes are further biased by in-context examples.
CVPR 2024, as first author.
Poster/OpenAccess/arXiv/Website/Code
We regularize dense prediction outputs in a shared task space, improving partially labeled multi-task learning while scaling linearly with task count.
CVPR 2021, as first author.
Video/OpenAccess/arXiv/Website/Code
We separate augmentation strategies for loss modeling and model learning, improving robustness to noisy labels, including a large gain on CIFAR-10 with 90% symmetric noise.
Get live translations for YouTube streams, crowdsourced from multilingual viewers.
HyperChat enhances your YouTube chat with a smoother, more feature-packed experience!
Capture YouTube chat messages based on filter rules alongside the standard chat, with persistent logs and import/export.
Officially commissioned by Cover Corp. and used as a yearly hololive English community event platform.
| Date | Status | Title | |
|---|---|---|---|
| Mar. 2026 | Draft | 2:11 | Bliss |
| Aug. 2023 | Demo | 1:59 | Celesta |
| Jan. 2023 | Done | 3:12 | Forfeit |
| Sept. 2022 | Done | 3:24 | Voyage |
| Sept. 2022 | Done | 3:53 | New Beginning |
| Jul. 2022 | Done | 1:58 | HyperChat Trailer Theme |
| Mar. 2020 | Demo | 1:17 | Dubstep Demo |
| Course | Title | |
|---|---|---|
| COMPSCI 2760 | Do AI Conferences' Ethics Reviews Steer Research Practices? (No!) | |
| COMPSCI 1050 | Does CS1050's Chatham House Policy Protect Against LLM-Powered Stylometric Linkage De‑Identification? (Yes!) | |
| COMPSCI 271 | Can Temporal Distance Maps Communicate Variability? (Yes!) A User Study with Maps of Transit Travel Times | |
| COMPSCI 175 | Implementing Portals in Unity | |
| HISTSCI 1990 | Browser Extension Standards: How Google Monopolized and Exploited the Web Browser Industry |
