Evaluate performance under equal compute budgets

Determine whether Temporal-Response Guidance retains its performance advantages over output-only on-policy distillation and direct alignment when methods are compared under equal training-compute budgets.

Background

Temporal-Response Guidance requires additional Jacobian-vector products, response-subspace construction, singular-value decomposition, and projection operations during training. The reported experiments compare methods using the same number of optimization updates rather than equal computational cost.

Consequently, the relative performance of Temporal-Response Guidance under matched compute budgets remains unresolved.