Gaussian Mean Field Variational Inference can Overestimate Predictive Variance

The paper isolates a specific failure mode: MFVI's underestimation of parameter uncertainty doesn't automatically translate to conservative predictive intervals. In high-density regions of the training data, the method can actually inflate confidence bounds, creating a blind spot where practitioners expect calibration but don't get it.

This connects directly to the uncertainty quantification benchmark from Argus (published same day), which tested 27 post-hoc calibration methods across vision-language models. That work flagged that uncertainty rankings collapse under distribution shift. This MFVI finding suggests part of the problem may be baked into the inference method itself, not just the post-hoc wrapper. If approximate Bayesian methods systematically misbehave in data-dense regions, then Argus's benchmark results for methods relying on variational inference may mask hidden failures in high-confidence zones. The two papers together imply that uncertainty evaluation needs to stress-test not just across models and datasets, but also across the geometry of the training distribution.

If follow-up work shows that practitioners using MFVI-based uncertainty in production regression systems (medical dosing, financial forecasting) have experienced miscalibration specifically on in-distribution predictions, that confirms this isn't just a theoretical curiosity. Watch whether the major Bayesian deep learning libraries (Pyro, Edward2) add explicit warnings or diagnostic tools for detecting this failure mode in the next 6-9 months.