Learning a Size-Weight Frontier for Synthetic-Augmented Inference
The paper proposes a coverage-tested way to use synthetic data without counting it as fully real.
Huang and Wang define a “size-weight frontier” that limits how many synthetic observations can be added at a given weight while still meeting target coverage. They estimate that frontier from related historical tasks and give a finite-sample guarantee for configurations at or below it. In experiments, LLM-generated responses helped augment opinion survey data while maintaining coverage and narrowing confidence intervals. ArXiv · AI/CL/LG's note
Huang and Wang define a “size-weight frontier” that limits how many synthetic observations can be added at a given weight while still meeting target coverage. They estimate that frontier from related historical tasks and give a finite-sample guarantee for configurations at or below it. In experiments, LLM-generated responses helped augment opinion survey data while maintaining coverage and narrowing confidence intervals. ArXiv · AI/CL/LG's note
score 4