Differentiable simulation is supposed to solve a fundamental problem in autonomous driving: models trained on open-loop data (predicting trajectories from frozen inputs) fail closed-loop (where predictions affect future states, and errors compound). Making the simulation differentiable lets gradients flow backward through time, teaching the model to recover from its own drift.
Yadav, Bohn, and Meisen (arXiv:2603.23393) found that the gradients leak something they shouldn't: the future. In a fully differentiable closed-loop simulator, loss gradients flow backward through the model's own previous predictions. These predictions were conditioned on ground truth state inputs. The backward flow creates a channel through which future ground truth information reaches past prediction steps, allowing the model to retroactively optimize earlier predictions using information it shouldn't have had.
The model learns to exploit this channel instead of learning genuine recovery behaviors. It doesn't learn to correct drift — it learns to make predictions that look good when evaluated against a future it has implicitly seen.
The fix is surgical: detach the computation graph between simulation steps. Break the gradient pathway that connects one timestep's output to the next timestep's input optimization. This forces the model to treat each step's incoming state as fixed — if it has drifted, it must recover from the drift rather than preventing it retroactively.
The detached version reduces target collisions by 33% compared to the fully differentiable approach. Breaking the gradient was the improvement.
The through-claim: differentiability through time creates an information channel that models exploit rather than learn from. The gradient that was supposed to teach recovery instead teaches retroactive optimization — learning to have been right rather than learning to get right.