Statistical Attribute Alignment for Black-Box Generative AI via Output-Level Post-Processing

Forum topic · 小凯 · 2026-09-29

Summary

A 2026 arXiv paper (2609.31607) by Kevin Jiang, Morgane Austern, Edgar Dobriban, and Jason M. Klusowski addresses statistical attribute alignment in black-box generative AI. The goal is to ensure the distribution of an attribute in AI-generated outputs matches a user-specified target, motivated by fairness (e.g., making protected attributes such as gender, race, or age follow desired distributions) and synthetic data generation (producing data representative of a target distribution). The authors study a practical black-box setting where users can repeatedly query a generative model, aiming to return m≥1 outputs whose joint attribute distribution is as close as possible to the target. They design algorithms for both exact and approximate alignment that minimize the expected number of generator queries, and prove optimality as the requested output count m approaches infinity. Experiments on text-to-image generation and profile generation with geographic locations show the post-processing algorithms significantly improve statistical attribute alignment, complementing prompt-based intervention methods.

Overview

Field: Machine Learning Authors: Kevin Jiang, Morgane Austern, Edgar Dobriban, Jason M. Klusowski Posted: 2026-09-25 arXiv: 2609.31607

Abstract

Generative AI systems are increasingly used, but aligning their outputs with user requirements poses a continuing challenge. This paper aims to ensure that the distribution of an attribute of an AI-generated output aligns with a user-specified target distribution.

Motivations:

Fairness: ensuring a protected attribute (e.g., gender, race, or age categories) follows a desired distribution
Synthetic data generation: producing generated data that is representative of a target distribution

Black-Box Setting

The authors study the practically important black-box access setting, where a user can repeatedly query a generative AI model. The goal is to return $m \ge 1$ outputs whose joint attribute distribution is as close as possible to the target distribution.

Contributions

Algorithms are designed for both exact alignment and approximate alignment scenarios.
The algorithms minimize the expected number of queries to the generator.
The authors prove the optimality of their algorithms as the requested number of outputs $m \to \infty$ .

Experiments

Experiments on text-to-image generation and profile generation with geographic location demonstrate that the proposed post-processing algorithms significantly improve statistical attribute alignment. They can serve as a complement to prompt-based intervention methods.

---

*Auto-collected on 2026-09-29*