Turning a vision-language model (VLM) into a robot brain has a hidden cost: it gets worse at understanding how far away things are.
Researchers probed a weight-matched pair of open-source models, Molmo2-ER and its action-trained counterpart MolmoAct2-LIBERO, to see how well each could decode depth information from every layer of its neural network. The action-trained version, a vision-language-action model (VLA), decoded depth worse than the base VLM at every single layer, a consistent gap the researchers call the floor. But the real surprise was in the final layers: the base VLM's depth understanding actually improves as information moves through its last layers, while the VLA's collapses instead, a sharper drop the paper calls the cliff. The team traced the cliff to interference from specific late-layer components (MLP writes) - turning those off recovered most of the lost depth signal, while disabling attention layers or making the same change in the base VLM did nothing comparable.
This matters because the whole promise of VLAs is reusing a VLM's world knowledge for physical tasks like grasping or navigating. If the process of teaching a model to act systematically degrades its spatial reasoning in a specific, localizable way, that is not a vague tradeoff - it is a targeted architectural failure mode that could explain why robot policies built on these models still struggle with basic geometry. It also hands the field a rare gift: a causal lever. Knowing exactly which layers and components erase depth understanding means future action-training methods could be built to protect it, rather than discovering the damage after deployment.
It is a reminder that bolting new capabilities onto a pretrained model rarely comes free - something upstream tends to break, even when the demo looks fine.