Statistical attribute alignment for black-box generative AI via output post-processing
Researchers propose query-efficient post-processing to align generative AI attributes with a target distribution.
The paper studies black-box generative models where a user can only issue queries and wants m outputs whose attribute distribution matches a specified target, such as a fairness constraint or a representative synthetic-data distribution. It presents algorithms for exact and approximate alignment that minimize the expected number of generator queries and proves they are optimal as m goes to infinity. Experiments on text-to-image generation and geocoded persona generation show the post-processing improves statistical attribute alignment beyond prompting.
- Black-box users query a generator and post-process its outputs.
- Algorithms minimize expected queries for exact and approximate alignment.
- Optimality is shown as the number of requested outputs grows.
- Tested on text-to-image and geocoded persona generation.
From the original article
1 min read at arxiv.orgGenerative AI systems are increasingly used, but aligning their outputs with user requirements poses a continuing challenge. Here, we aim to ensure that the distribution of an attribute of an AI-generated output aligns with a user-specified target. This is motivated by examples such as fairness, where we want to ensure that a protected attribute (e.
By Kevin Jiang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.