Spectra
Test-time prior adaptation for diffusion-based simulation-based inference, without additional simulation or training.
- Question
- Amortized simulation-based inference models are trained under one prior, but scientific analyses often need revised priors as knowledge accumulates or alternative assumptions are tested. Can a pretrained model be reused under a new prior without retraining?
- Built
- A test-time adaptation method that uses an exact score-transport identity to obtain the adapted score from a frozen diffusion model in closed form for structured prior changes, together with a way to combine posterior components for mixture priors.
- Result
- Across six SBI benchmarks, accurate adaptation under strong prior shifts at low online sampling cost.