Tomoya Wakayama

2 accepted papers

2026

A Decision-Theoretic View of Test-Time Training: When, How Far, and Which Directions to Adapt

ICML 2026poster

Test-time training (TTT) adapts a pretrained model to each prompt via parameter updates, improving accuracy under pretraining-to-test distribution shifts. Yet, its performance often suffers from instability and sensitivity to hyperparameters such as update steps and subspace. We explain this behavio…

Cited by 0 Source ↗Scholar ↗

2026

In-Context Learning Is Provably Bayesian Inference: A Generalization Theory for Meta-Learning

Tomoya Wakayama, Taiji Suzuki

ICML 2026poster

This paper develops a finite-sample statistical theory for in-context learning (ICL), analyzed within a meta-learning framework that accommodates mixtures of diverse task types. We leverage a Bayes risk identity that separates the total ICL risk into two orthogonal components: Bayes Gap and Posterio…

Cited by 0 Source ↗Scholar ↗