Beyond Expected Information Gain: Stable Bayesian Optimal Experimental Design with Integral Probability Metrics and Plug-and-Play Extensions

The deeper issue here is not just numerical stability: KL-divergence's sensitivity to rare events means that experimental designs optimized under it can be systematically skewed toward avoiding low-probability outcomes rather than genuinely maximizing information, which quietly corrupts the entire acquisition strategy. The plug-and-play framing also suggests the authors are targeting practitioners who want to swap objectives without rebuilding inference pipelines.

The information-gain thread running through recent coverage is worth noting. The IG-Search paper from mid-April used step-level information gain signals to guide LLM search behavior, which is a different domain but shares the same foundational tension: information gain metrics are theoretically appealing but practically fragile when probability distributions have thin tails or rare support. That paper worked around the problem by operating on retrieved document confidence rather than raw KL quantities. This new work attacks the fragility more directly at the mathematical level. The connection is not tight, but both papers are essentially arguing that naive information gain objectives break in practice and need structural fixes.

The real test is whether the proposed metrics hold up in high-dimensional posterior settings common in drug discovery or materials science experimental loops, where rare-event bias is most damaging. If independent groups replicate the stability gains on real-world closed-loop design benchmarks within the next year, the case for retiring KL-divergence as the default becomes hard to ignore.