- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper introduces a training-free generation framework for time-series modeling. For reward-tilted sampling, the authors employ a stochastic optimal control formulation, where naive reward guidance has pitfalls, and propose an inference-time control method based on the adjoint matching idea. For a control parameterization appropriate for time-series modeling, they restrict the control space to low-frequency Fourier or coarse wavelet subspaces.
Critical strengths
- For time-series modeling, restricting the control space to Fourier and wavelet subspaces for better control parameterization is interesting.
- The proposed method shows superior results across multiple datasets and constraint types compared to the baselines.
Critical weaknesses
-
The motivation for functional parameterization in Section 4.1 is interesting, but it is unclear whether the proposed parameterization, which restricts the control space, yields a sufficiently accurate approximation to the optimal control. Specifically, the bound in Theorem 4.1 follows immediately from Girsanov’s theorem and the data-processing inequality. It is expressed entirely through the distance between the optimal control and its projection onto the restricted Fourier or wavelet subspace. The approximation bias therefore depends on how much of the optimal control lies outside this subspace, and that is exactly what the theorem leaves open. The bound vanishes only when the optimal control already lies in the subspace. If a substantial part of it lies outside, the bound becomes large and tells us nothing.
-
In my view, the novelty lies mainly in Section 4.2, since Section 3 largely builds on standard results in the stochastic optimal control literature, and the Fourier and wavelet restrictions in Section 4.1 are fixed linear filters applied to the control. Section 4.2 borrows the adjoint idea from Adjoint Matching. It restricts the control to low-frequency or coarse wavelet components, then updates it iteratively by backpropagating the terminal loss through the sampler. Some methods estimate the h-function in Doob’s h-transform without iterative backpropagation, as described in [1, Section 3.2]. I wonder how the proposed method compares with these methods in terms of performance.
-
It is unclear how the proposed algorithm recovers the optimal tilt. The paper states that the intractable score term (Doob’s h-transform function) is bypassed via PMP (lines 286–287). However, the adjoint used to define the optimal control remains the same intractable object (a conditional expectation over future noise [2,3]). In [2,3], a deterministic adjoint ODE such as Eq. (13) is sufficient because the per-path adjoint serves as a regression target. Its conditional expectation given the intermediate state yields the true adjoint, so the loss gradient is unbiased and the fixed point of the optimal control problem is preserved. However, Algorithm 1 seems to have no such step. The per-path adjoint omits the stochastic term and may not represent the required conditional expectation, yet it is used directly to form the control in the single-pass setting or update it during refinement. It is therefore unclear whether Algorithm 1 preserves the fixed point corresponding to the optimal control.
[1] Potaptchik et al., Meta Flow Maps enable scalable reward alignment
[2] Domingo-Enrich et al., Adjoint Matching through the Lens of the Stochastic Maximum Principle in Optimal Control
[3] Park et al., Functional Adjoint Sampler
Questions for authors
-
How is the terminal loss function L is defined for the experiment? I wonder how the gradient is computed for the reward-guidance or backpropagation.
-
Minor point: The Schrödinger bridge problem targets minimize the KL divergence to the reference process with the marginals prescribed at both ends. However, Eq.(4) fixes only the initial distribution and replaces the terminal-marginal constraint with a soft terminal cost. It is therefore a KL-regularized optimal control problem with a terminal cost, not a bridge between two given marginals.