Daily AI Paper

Generative AI can make impressive outputs, but it still struggles to match the statistical properties users actually want. This paper tackles a practical version of that problem: given a black-box generator, how can we make the outputs follow a target distribution for some attribute, like gender balance in a dataset or demographic proportions in synthetic personas, without retraining the model? The authors propose output post-processing methods that repeatedly query the generator and then select or adjust the returned samples so the final set aligns with the desired attribute distribution as closely as possible. A key strength is that the approach works even when you can’t inspect or fine-tune the model internals. They also prove the methods are query-efficient and optimal in an asymptotic sense. In experiments on text-to-image generation and geocoded personas, the algorithms improve alignment beyond prompting alone. That matters because it offers a lightweight way to steer real-world generative systems toward fairness, representativeness, and user-defined constraints.