A new arXiv preprint asks when feedback actually helps in learned experimental design. The authors audit a learned scientific controller on two synthetic voltage-clamp tasks and find that it can outperform a misspecified planner without benefiting from feedback at all.
The audit compares a training-selected fixed protocol, current masking, and a Bayesian reference. The abstract is truncated in the source, so full details of the comparison are not yet available.
The result is a caution for researchers who assume closed-loop feedback always improves performance. Even when a controller wins, the win may come from training or design choices rather than from using feedback.