A Decision-Theoretic View of Test-Time Training: When, How Far, and Which Directions to Adapt
Abstract
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 eigen-directions. This perspective underpins the following results: (1) we show when fixed update steps and subspaces fail under distribution shifts, motivating adaptive strategies; (2) we prove that selecting update steps via prompt evidence admits a PAC-Bayes guarantee against overfitting; and (3) we characterize the Bayes-optimal update subspace under a linear-Gaussian correction model, yielding a scoring rule for selecting Transformer blocks and heads. Our theory helps explain the empirical instability of TTT, taking a step toward principled guidance for when, how far, and which directions to adapt.
Related
- A Decision-Theoretic Framework for Comparing Likelihood Ratio Methods for the Rare Type Match Problem
- A decision-theoretic framework for uncertainty quantification in epidemiological modelling
- Decision-theoretic recruitment and validation planning for depression surveys: An illustrative Chinese application
- Decision-Aware Training for Sample-Based Generative Models
- Certifying Whether to Adapt: Baseline-Relative PAC-Bayes-MMD Guarantees for Test-Time Adaptation
BibTeX
@misc{wakayama2026decisiontheoretic,
title = {{A Decision-Theoretic View of Test-Time Training: When, How Far, and Which Directions to Adapt}},
author = {Tomoya Wakayama},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2606.15569},
eprint = {2606.15569},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2606.15569},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - A Decision-Theoretic View of Test-Time Training: When, How Far, and Which Directions to Adapt
AU - Tomoya Wakayama
PY - 2026
DA - 2026/10/19
PB - arXiv
DO - 10.48550/arxiv.2606.15569
UR - https://arxiv.org/abs/2606.15569
AB - 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 eigen-directions. This perspective underpins the following results: (1) we show when fixed update steps and subspaces fail under distribution shifts, motivating adaptive strategies; (2) we prove that selecting update steps via prompt evidence admits a PAC-Bayes guarantee against overfitting; and (3) we characterize the Bayes-optimal update subspace under a linear-Gaussian correction model, yielding a scoring rule for selecting Transformer blocks and heads. Our theory helps explain the empirical instability of TTT, taking a step toward principled guidance for when, how far, and which directions to adapt.
ER -