Double Preconditioning (DoPr): Optimizing for Test-Time Rollout, Not Validation Loss

arXiv·medium signal

DoPr targets 'test-time feedback' — the documented mismatch where models trained with one-step prediction loss (cross-entropy, L2) are deployed by rolling out along their own predictions, as in autoregressive generation, flow models, and robot policies. It proposes a double-preconditioning optimization scheme that directly improves downstream rollout metrics rather than validation loss. A useful lens for anyone whose eval metric diverges from training loss.