The day in conversation for 2026-10-23

The discussion: Tom: I’m really focused on that gradient discrepancy characterization mentioned in the entropy-regularised objectives paper. It seems like defining a computable measure for suboptimality is the bottleneck because standard divergence metrics don't capture the specific geometric structure of kernel Stein discrepancy as well as this approach does. Jane: That computational feasibility is key, but I’m more concerned about how that characterization interacts with selection methods. If we have a measure quantifying how far an objective function is from optimal, how do we reliably use that metric to prune or select the best model among a set when those models are drawn from different priors? Lu: That leads into the OptCS work. Their framework seems designed to handle data-dependent scores, but I wonder if the gradient discrepancy we’re discussing needs an assumption about the smoothness of the underlying kernel function for that characterization to remain tractable. What happens if we deal with highly non-smooth objectives in our variational setup? Meng: If the objective is non-smooth, perhaps we should pivot to something more robust than direct gradient comparison, maybe incorporating a robust loss function estimate instead of relying purely on the discrepancy measure. Did anyone look at how FRESH handles label noise—does its framework for combining patient-level data with population summaries account for that inherent uncertainty in the