BRIDGE: Turning Frozen Foundation-Model Features into Explicit Structure

Pretrained generative models hold information that is useful for tasks they were never trained to solve. Getting at it is the hard part: a downstream model has to select the relevant intermediate features and still respect the dependencies among its own outputs. We present BRIDGE, a lightweight architecture that reads the intermediate activations of a frozen model and turns them into the different representation a task requires. A small set of capture sites is compressed into a queryable state of fixed size. Each output element forms its own query from its current state and geometry, a shared global state lets elements exchange information before they are updated, and the update itself is whatever the task needs. The reader is indifferent to which model supplied the activations and to the form of the output head. We instantiate BRIDGE for sparse-view 3D surface completion from a frozen video model: it predicts a fixed population of oriented surface points and carries their positions and normals jointly with conditional flow matching, so that each local prediction sees its own geometry and scene-level information together. The same reader also runs in the opposite direction. Trained on a frozen vision–language model’s activations and then frozen, it becomes a differentiable geometric critic whose error is backpropagated into a low-rank adapter of the language model. With the critic removed, the adapted model raises its strict successes on held-out objects it had failed from 10 to between 17 and 20 of 60 under an unchanged criterion, while an untrained critic of the same architecture yields no gain. A reader of intermediate representations can thus both extract structure from a frozen model and feed structure back into it.