A new method tracks how training targets change predictions

If changing a training label or value could alter a model’s predictions, how can we estimate that change without retraining? Existing attribution measures how predictions respond when examples receive different loss weights.

The unresolved issue is whether that weight response can reliably stand in for changing the targets themselves.

The target-response Jacobian estimates how each training target affects model predictions, answering a question that weight-based attribution does not measure.

The researchers trained models, changed training targets in sampled ways, and refit each model from its original fitted parameters. They compared each method’s predicted change with the corresponding refit change, scoring each prediction across target changes and then taking the median across evaluation predictions. A negative score means estimates were worse than predicting the average refit change.

On three tabular datasets, the target-response Jacobian closely matched refit shifts, while influence scored below zero. This supports the claim that weight sensitivity was a poor substitute for target sensitivity in these tests.

The target-response Jacobian estimates a first-order response at a fitted stationary point, so accuracy for larger target changes depends on how nonlinear the refit is.

The method requires access to model derivatives and Hessian solves, and the reported results cover specific models, target changes, and evaluation settings.