I research methods that make large language models perform better by spending more compute at inference time, an approach known as inference-time scaling.
- First-author papers
- ICML 2026, ICLR 2026, NeurIPS 2025 (Spotlight)
- Before Sakana AI
- Apple (Japanese Siri), Recruit
- PhD
- Theoretical Physics, The University of Tokyo
Selected Publications
-
We propose an inference-time scaling method for masked diffusion language models in which multiple models collaborate during unmasking, with Monte Carlo tree search choosing among their deterministic unmasking steps. It outperforms existing inference-time scaling methods on coding benchmarks and scales well on math reasoning.
-
LLMs often fail to make choices according to given probabilities: asked to flip a fair coin, they don't return heads and tails equally often. We show that prompting alone largely fixes this. Asking the model to first write a random string and then derive its answer from it brings its choices close to the target probabilities and makes open-ended answers more diverse.
-
We propose AB-MCTS, an inference-time scaling method for LLMs that extends Monte Carlo tree search: using Bayesian reward modeling, it decides at each step whether to sample a new answer or refine an existing one. Across a range of coding and engineering tasks, it consistently outperforms repeated sampling and standard MCTS. As part of this work, I developed TreeQuest, the accompanying open-source library.
Other Publications
Full list on Google Scholar.