Source
Zhang, Zhang, Zardini, Amin and Ozdaglar (2026), LIT-490 — ARXIV-2608.29507 — read as NOTE-239.
What to do
The reordering is LIT-490’s, which proposes it on the reading that the denoiser is an approximate projection onto the data support, proves finite-time guarantees for it in three geometries, and shows in double-integrator and unicycle tracking that the standard ordering’s plans lose much of their advantage once rolled out through the true dynamics while the reordered ones do not.
At each reverse step you have a noisy iterate, a pretrained denoiser, and a
differentiable objective. The standard implementation denoises, then adds the
gradient correction outside. Reverse that: apply −η∇f to the noisy
iterate, then call the denoiser on the result, then take the usual DDIM step.
One gradient evaluation, one denoiser call, one trajectory. No score Jacobians, no fine-tuning, no auxiliary model, no repeated batch generation — which is what the nearest optimization-oriented alternatives require.
Why the order is not a detail
If the data lie on a structured feasible set — a manifold, a convex region, a family of dynamically admissible trajectories — the objective gradient need not point along it. Added after denoising, the gradient is the last thing that happens: you end the step off the geometry, and nothing corrects it. Added before, the denoiser is the correction, because it maps noisy points back toward the data support.
THEORY-053 is the account — the Stein denoiser as an approximate
projection — and with it the reverse process is an inexact projected-gradient
method with finite-time guarantees on linear subspaces, compact convex sets
and compact Riemannian submanifolds. It is Proposed, and the practice does
not depend on the proofs: the empirical failure of the other ordering is
direct.
What it buys, and where the evidence actually is
The trajectory-planning experiments are the informative ones, and the result is about measurement as much as optimization. In double-integrator and unicycle reference tracking, the standard ordering produces plans that track the reference closely — and much of that advantage disappears once the controls are rolled out through the true dynamics. The plans were not dynamically consistent; they were fitted to the reference. With the gradient inside the denoiser, planned and executed trajectories nearly coincide, and it reports the lowest rollout cost and dynamic-feasibility error among guided methods on both random references.
So the honest statement of the benefit is not “better objective values”. It is that the standard ordering optimizes the plan and this one optimizes something you can act on — and a practitioner comparing the two on planned cost alone would pick the wrong one.
Conditions
The guarantees describe an idealization. The proofs assume an exact Stein posterior-mean denoiser and that the learned support faithfully represents the feasible set. What you run is a trained network approximating that denoiser, on a model whose support may include infeasible regions. The authors name misspecification as future work, and that gap is the whole distance between the theorem and the deployment.
The setting is narrow. Deterministic DDIM, variance-exploding diffusion, an exponentially decaying noise schedule. Other samplers and schedules “require tracking their time-varying relaxation and denoising errors” — the theory does not transfer as written, though the ordering argument is cheap enough to test wherever you are.
The reinforcement-learning result is the weakest arm and the most citable one. On Hopper, Walker and HalfCheetah (medium-v2) under a Diffuser-style pipeline, this achieves “slightly higher” normalized returns — the authors’ own word — against one baseline on three tasks, with guidance scale tuned separately per method. D4RL numbers travel further than a unicycle study; here the unicycle study is the better evidence.
It is not free of the tuning problem. A guidance stepsize still has to be chosen, and the comparison tunes it separately for each method, which is correct practice and means the reported gap is between two tuned systems rather than a drop-in improvement.
Your feasible set is whatever the training data said it was. Projection onto the learned support is a feasibility guarantee only insofar as the training trajectories were themselves feasible.