Distribution Shift
Developing TILT, a one-step adaptation method that avoids explicit density-ratio estimation, with theoretical equivalence to relative density reweighting and improved stability under covariate shift.
MIT EECS · LIDS · IDSS
Kakei Yamamoto
山本かけい
PhD student in Computer Science at MIT, advised by Prof. Martin J. Wainwright. I study robust learning and inference, with interests in post-training, distribution shift, and LLM generalization.
Research Assistant, Laboratory for Information & Decision Systems
Profile
Biography
I am a PhD student in Computer Science at MIT EECS and a research assistant at LIDS, advised by Martin J. Wainwright. I am also affiliated with IDSS.
My research focuses on developing robust learning and inference methods, particularly for post-training and adaptation under distribution shift. I am also interested in extending the ability of large language models to generalize to settings not represented in their training data.
Before MIT, I worked on deep learning theory at RIKEN AIP and studied Applied Mathematics and Computer Science at the University of Tokyo under Taiji Suzuki.
Interests
- Inference
- Fine-tuning
- Generative models
- Machine learning theory
- Optimization
Education
- PhD in Computer Science Massachusetts Institute of Technology · Sep 2023–May 2028 expected
- MS in Computer Science Massachusetts Institute of Technology · Sep 2023–Feb 2026
- BEng in Applied Mathematics and Computer Science The University of Tokyo · Apr 2019–Mar 2023
Awards
- Takenaka Overseas Scholarship Takenaka Scholarship Foundation · 2023–2028; tuition and living support
- Kiyo Sakaguchi Scholarship The Prudential Life Insurance Company, Ltd. · Selected 2023; declined monetary support due to concurrent Takenaka scholarship
Research
Current directions and foundations from earlier work.
Model Distillation
Studying functional gradient descent for student-teacher estimation. Residual-as-Teacher learns from residuals to mitigate teacher-bias propagation under high-dimensional covariate shift.
LLM Evaluation
Sakana AI · Jul–Aug 2026 internship
Calibrating black-box judges with counterfactual probes in PRISM, and studying debiasing and robust inference for AI agents.
Mean-Field Optimization
Proving convergence and global optimality for feature-learning actor-critic methods, and analyzing Langevin dynamics for distributional minimax optimization.
Spiking Neural Networks
Developing timing-based backpropagation for multi-spike and recurrent networks, with earlier work on sparse-firing regularization and efficient neural computation.
Publications & Talks
2026
PRISM: Calibrating Black-Box Judges with Counterfactual Probes
arXiv, 2026
Uses responses to counterfactual probes and human labels to calibrate scores from black-box LLM judges.
2026
TILT: Target-induced loss tilting under covariate shift
NeurIPS, 2026
Introduces Target-Induced Loss Tilting, a one-step covariate-shift adaptation method that decomposes the source predictor into a deployed predictor and a target-penalized auxiliary component, implicitly inducing relative importance weighting without estimating density ratios.
2026
Residual-as-Teacher: Mitigating Bias Propagation in Student-Teacher Estimation
arXiv, 2026 · Annals of Statistics (under review)
Proposes a functional gradient descent framework for model distillation, learning from residuals to mitigate teacher-bias propagation and outperform standard pseudo-labeling under high-dimensional covariate shift.
To appear
Tilted gradient descent
2024
Mean Field Langevin Actor-Critic: Faster Convergence and Global Optimality beyond Lazy Learning
ICML, 2024
Develops mean-field Langevin actor-critic methods for feature-learning policy optimization.
2024
Symmetric Mean-field Langevin Dynamics for Distributional Minimax Problems
ICLR Spotlight (top 6%), 2024
Extends mean-field Langevin dynamics to minimax optimization over probability distributions.
2024
Can Timing-Based Backpropagation Overcome Single-Spike Restrictions in Spiking Neural Networks?
IJCNN, 2024
Studies timing-based backpropagation beyond single-spike restrictions in spiking neural networks.
2023
Sparse-firing regularization methods for spiking neural networks with time-to-first-spike coding
Scientific Reports, 2023
Introduces spike-timing-based regularization methods to reduce firing frequency in TTFS-coded spiking neural networks.
2022
Proposal of a Numerical Method for Calculating the Gamma Function by the Double Exponential Formula
JSIAM Annual Meeting, 2022
Presented an improved double-exponential quadrature method for computing the gamma function.
Portfolio
Selected Works
Calendar app · Since 2023
Longitude°
First built as a personal iPhone beta in 2023. One tap switches the entire calendar; Dual Time aligns one day to two local clocks. Now on iPhone, with Android closed-test participants wanted.
Released web game · 2026
BiteType! and BiteType! Classroom
Created and released a browser typing game set in a burger stand, plus a classroom mode for no-login school practice, teacher links, class results, and guided typing settings.
Community leadership · 2026
Japanese Association of MIT
Serving as JAM President. Led Hanami Festival as the representative organizer, planned Beneath the Great Wave from zero, and designed and built the JAM website.
Service
Leadership
Japanese Association of MIT (JAM)
President · 2025-
Representative organizer for Hanami Festival 2026; originated and planned Beneath the Great Wave; built the organization website and programming infrastructure.
BiteType!
Creator and developer · 2026
Designed, implemented, and released the typing game and its classroom product, including gameplay, audio, analytics, teacher controls, and Cloudflare deployment.
Teiyukai
Executive board member, General Affairs Director · 2022-2023
Served on the executive board of the University of Tokyo engineering student alumni/community organization.
UTokyo Philharmonic Orchestra
President · 2020-2021
Led the student orchestra while playing trumpet.