A new audit finds that routing entropy, the internal confusion level of a dynamic neural network, mostly isn't a useful warning sign beyond what the model's confidence score already tells you.
Researchers tested two image classifiers, Attention-Residual versions of Swin-Tiny and DeiT-Small, trained from scratch on CIFAR-10 and CIFAR-100. Each model produces a routing trace alongside its prediction, a record of how its internal pathways activated, and the assumption in some uncertainty research is that a scattered, high-entropy trace flags a shaky answer. The team ran three checks: does the signal show up at a fixed confidence level, does it hold up across repeated training runs with different random seeds, and can a separate predictor actually extract useful information from it. Across 24 paired training runs, none of the 30 statistical tests survived correction for multiple comparisons, and the one borderline positive result did not reappear when the run was repeated with a different seed.
The interesting part is what the routing trace could do: it beat a shuffled, meaningless version of itself at predicting correctness, but it still lost to the model's own confidence score on both standard scoring methods. That is the distinction worth flagging - beating noise is not the same as beating the baseline you already have. The team also measured how large a real effect would need to be before their test could catch it, and found their method would miss a meaningful share of a true signal even in the best case, below the bar they had set for trusting it on real labels.
The paper does not declare routing entropy worthless, only that this detector was not sharp enough to prove otherwise - worth remembering next time a demo claims to read a model's mind by watching which internal pathways light up.