Risk-Averse Inference-Time Diffusion Model Alignment with Finite-Sample Guarantees
Abstract
Inference-time alignment steers diffusion models toward downstream objectives without costly fine-tuning, but existing methods typically optimize expected cost (or reward) without explicitly controlling tail risk. In this work, we propose a risk-averse framework for inference-time alignment of diffusion models that balances expected cost and Conditional Value-at-Risk (CVaR) under trajectory-level Kullback-Leibler regularization. We characterize the optimal path distribution and reverse transitions as functions of the CVaR threshold, reducing the alignment problem to a one-dimensional optimization problem over the threshold. We show that this one-dimensional optimization problem can be nonconvex in general, establish sufficient conditions for convexity, and introduce quadratic regularization to enable efficient derivative-free ternary search. Under suitable assumptions, we provide finite-sample guarantees for threshold optimization and value estimation, and combine them into a terminal total-variation bound relative to the regularized target under exact resampling. Experiments on synthetic data and image generation using Stable Diffusion demonstrate improved tail performance over risk-neutral baselines while maintaining competitive average cost.
Related
- Risk-Averse Bayesian Games with an Unknown Type Distribution: Bayesian Learning, Equilibrium Analysis, and Finite-Sample Guarantees
- Finite-Sample Guarantees for Fully Discrete Conditional Diffusion Models
- Risk-Averse Decision Making with Multi-Level Reliability Guarantees
- Steepest Guidance: A Practical and Principled Approach to Inference-Time Alignment of Flow and Diffusion-based Models
- Diffusion Flow Matching: Dimension-Improved KL Bounds and Wasserstein Guarantees
BibTeX
@misc{liang2026riskaverse,
title = {{Risk-Averse Inference-Time Diffusion Model Alignment with Finite-Sample Guarantees}},
author = {Yuchen Liang and Xian Yu},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.17579},
eprint = {2610.17579},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.17579},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Risk-Averse Inference-Time Diffusion Model Alignment with Finite-Sample Guarantees
AU - Yuchen Liang
AU - Xian Yu
PY - 2026
DA - 2026/10/13
PB - arXiv
DO - 10.48550/arxiv.2610.17579
UR - https://arxiv.org/abs/2610.17579
AB - Inference-time alignment steers diffusion models toward downstream objectives without costly fine-tuning, but existing methods typically optimize expected cost (or reward) without explicitly controlling tail risk. In this work, we propose a risk-averse framework for inference-time alignment of diffusion models that balances expected cost and Conditional Value-at-Risk (CVaR) under trajectory-level Kullback-Leibler regularization. We characterize the optimal path distribution and reverse transitions as functions of the CVaR threshold, reducing the alignment problem to a one-dimensional optimization problem over the threshold. We show that this one-dimensional optimization problem can be nonconvex in general, establish sufficient conditions for convexity, and introduce quadratic regularization to enable efficient derivative-free ternary search. Under suitable assumptions, we provide finite-sample guarantees for threshold optimization and value estimation, and combine them into a terminal total-variation bound relative to the regularized target under exact resampling. Experiments on synthetic data and image generation using Stable Diffusion demonstrate improved tail performance over risk-neutral baselines while maintaining competitive average cost.
ER -