Reward optimization and few-step distillation often optimize different objects. This work asks whether both can be formulated around the finite-step operator actually used at inference.
The idea
A corrective transport field acts alongside a frozen base generator. The solver rule stays fixed, while the correction is conditioned on the evaluation budget. Terminal improvement provides a signal for pathwise optimization, and a conditional distribution-matching objective connects the result to target data.
The research covers conditional generation and restoration. It combines reinforcement learning, distillation, and finite-step optimization.
Reading the demonstration
The experiment below isolates one idea: changing a vector field changes a finite-step transport. It uses an explicitly specified two-dimensional correction, not a trained network. It does not reproduce the paper’s optimization procedure or report research performance.
Interactive toy experiment
A correction changes the path
Adjust the correction strength and Euler budget. Both paths start from the same particles and share a fixed base field.
36 particles · 8 Euler steps · RMS endpoint distance to analytic reference: 0.0096. At zero correction, both endpoints coincide.
How this toy works
b = (1, 0); m(t) = (−1.2 + 2.4t, 0.6t); r = (1.4, 0.6) − 0.8(z − m). Euler update: z ← z + (b + αr)/N. At α = 1 the analytic endpoint is m(1) + exp(−0.8)(z₀ − m(0)).
The corrective field is hand-specified, not learned. Endpoint distance measures this toy system, not generation quality or the paper’s optimization results.
Synthetic illustration · Not a trained-model output or a published benchmark.