When generative AI is guessing: our AISTATS 2026 paper on epistemic uncertainty in diffusion models
Our paper “Quantifying Epistemic Uncertainty in Diffusion Models” (with Aditi Gupta, Raphael A. Meyer, Yotam Yaniv, and N. Benjamin Erichson) has been accepted at AISTATS 2026.
Diffusion models are random by design: every run starts from fresh noise and yields a different sample. That randomness is a feature, but it hides a problem — when an output looks strange, the model may be exploring a genuinely unusual possibility, or it may be guessing in territory it never learned, and total variance across samples cannot tell the two apart. Our method, FLARE, isolates the epistemic part: the variation that comes from the model not knowing. It builds a Laplace approximation over a random subset of parameters drawn from every layer of the network and propagates the resulting uncertainty through the full denoising trajectory, with the relative error provably shrinking as $O(1/\sqrt{m})$ in the number of sampled parameters. The common last-layer shortcut carries no such guarantee, and in our experiments it misses exactly the samples one would most want to catch. Used as a filter on generated time series, FLARE closed up to 93 percent of the gap between generated and real data, where the last-layer approximation sometimes ranked samples backwards.
🔗 Links: