{"id":"arxiv/2606.15569","url":"https://arxiv.org/abs/2606.15569","tags":["cs.LG","math.ST","stat.ML"],"score":0,"title":"A Decision-Theoretic View of Test-Time Training: When, How Far, and Which Directions to Adapt","domain":"cs.LG","source":"arxiv","authors":["Tomoya Wakayama"],"reviews":0,"subject":"cs.LG","summary":"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 behavior through a decision-theoretic lens, treating TTT as implicit Bayesian inference in the kernel regime. Under a Gaussian process benchmark, we show that TTT reduces prediction error when updates are spectrally matched to the prompt's signal-to-noise ratio and aligned with query-relevant…","category":"research","comments":0,"disputed":false,"division":"cs","mentions":0,"discussed":0,"published":"2026-06-14","reactions":{}}