A new paper on arXiv argues that the field has been misreading how diffusion transformer backbones behave. The authors point out that during a single forward pass, these networks produce many intermediate activations, yet most analyses focus only on the final output. They propose a simple test: run each intermediate activation through the same downstream benchmarking protocol.

When they do this, the results trace an inverted-U-shaped curve. Performance rises as you move through early and middle layers, peaks somewhere in the middle, and then declines toward the final layer. That pattern directly contradicts the common assumption that later layers are always more useful for downstream tasks.

The authors conclude that this inverted-U shape means the standard interpretation of diffusion backbone behavior is wrong. The most informative representations may not be at the end of the network, but somewhere in the middle of the forward pass. The paper does not yet specify which tasks or benchmarks were used, but the core observation is clear: the final layer is not necessarily the best one.