Wittawat Jitkrittum
About me
My name is Wittawat Jitkrittum (วิทวัส จิตกฤตธรรม) (). People call me "Witta". I am a research scientist working on LLMs at Google DeepMind, New York. My core work is to design good objective functions for improving LLM training efficiency, and inference efficiency. Some research questions and example projects I worked on are as follows.
-
Model routing for efficient LLM inference
With multiple distinct language models in the serving pool, can we intelligently route each given input query to the right model in the pool?
Using small models as much as possible, and reserving large models only for few complex queries will save inference cost. Relevant works: Jitkrittum et al., 2025, Jitkrittum et al., 2023, Rabanser et al., 2025, Gupta et al., 2024, and Li et al., 2025 (for routing to text-to-image models). -
Teacher-based LLM training
Can we leverage an already trained language model to train another one?
Wang et al., 2024 explores ways to train a small model only on tokens deemed easy according to a large teacher model. Rawat et al., 2024 explores the use of a small teacher model to help train a large student model. -
Model comparison (2016-2023)
Looking only at samples, are two multivariate distributions the same?
We developed sample-efficient statistical tests (Jitkrittum et al., 2016, Kübler et al., 2022) for this problem. When it is a comparison between a sample and a known model, these kernel-based tests can be used: Jitkrittum et al., 2017, Jitkrittum et al., 2020, Muandet et al., 2020, Kanagawa et al., 2023. It is often more practical to decide which of \(K\) candidate models better fit the data. See Jitkrittum et al., 2018 (\(K=2\)) and Lim et al., 2019 (\(K > 2\)).
I also worked on image generative modeling in the past. Please see the full list of publications. My talks can be found here.
Online presence
ArXiv DBLP Github Google Scholar LinkedIn Orcid ID ResearchGate Semantic Scholar Twitter
Some highlights
My resume in PDF can be found here
Experience |
|
Education |
|
Filed patents |
|
Selected awards |
|
Selected academic services |
|
Co-organizer of regional events |
|
Last update: 2025-09-09