- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes HyRef-SB, a framework for conditional stochastic transport on a hyperbolic–Euclidean product manifold. An invertible calibration maps preprocessed observations into geometric coordinates. An endpoint-preserving teacher supplies trajectories whose velocities train a reference drift. This reference is then frozen while a residual controller is optimized through stochastic rollouts, using intrinsic control energy, soft terminal distribution discrepancies, and optional support costs.
Critical strengths
-
The same metric defines diffusion, control energy, and exponential-map updates. The factors associated with the Poincaré convention are handled consistently: the radial coefficient, transverse amplification, and origin-matched Euclidean comparison agree with this convention. Preserving all preprocessed coordinates also avoids conflating geometric calibration with dimensionality reduction.
-
The comparison drift $ b^\dagger=b_\ast+\frac{\sigma^2}{2}\nabla_g\log p_t $ correctly compensates for diffusion while preserving the teacher marginals. The resulting bound separates reference regression error from the density correction required by stochastic dynamics. The paper appropriately retains the Fisher-information term even when regression error vanishes.
Critical weaknesses
-
Theorem 5.4 concerns approximate minimizers over population path measures as the endpoint penalty tends to infinity. The implementation uses a finite penalty, restricted neural controls, empirical discrepancies, and a fixed discretization. Corollary C.14 correctly identifies approximation, optimization, and estimation errors, but none of these errors is controlled experimentally.
-
All main benchmarks and geometry/module ablations use a single training seed. The five-seed diagnostic is useful but does not establish robustness of the navigation, tree, or cellular rankings. This matters especially for small differences, such as Mouse Hematopoiesis intermediate Wasserstein errors.
-
Appendix F.5.1 specifies that Clonidine and Trametinib use all 50 PCs for MMD but only the first two PCs for and . This distinction is not prominent in Tables 3 and 7 or their interpretation. The paper should report full-dimensional Wasserstein or sliced-Wasserstein diagnostics alongside the two-PC results and identify evaluation coordinates directly in the table captions.
-
Equation (25) combines generated-endpoint readouts with fixed source-conditioned weights . However, the construction of these weights for unseen source cells is insufficiently specified.
-
Collision detection examines only interior grid points. This is a valid discrete evaluation, but it can miss obstacle crossings between observations. The reported near-perfect safety should therefore be supported by grid-refinement or more conservative crossing checks, with the total number of evaluated paths stated.
-
The existing geometry and Brownian-reference ablations are helpful, but they do not establish how much performance comes from residual stochastic optimization beyond the teacher and fitted reference. Cellular experiments also use a small total noise scale of . Comparisons against the frozen reference, a deterministic counterpart with matched architecture and action penalty, and several noise/control-weight settings would help establish when the full stochastic bridge formulation is necessary.
Notation is generally consistent, but is reused for teacher parameters, the endpoint-penalty multiplier, and the calibration coefficient in Appendix F.3. Distinct symbols would improve readability. Figure 2 is too small for reliable inspection of its axes and trajectories. A compact notation table and a single implementation specification would be more useful than some of the repeated explanations of standard identities.
Questions for authors
-
What are the exact numerical safeguard maps and thresholds? Are they identical for controlled and reference transitions as required by Proposition C.19, and how often are they activated?
-
Can you evaluate endpoint discrepancy and control energy across increasing penalty strengths on a problem with a known bridge solution? Which parts of the population penalty theorem are empirically supported by the trained model?
-
Do the main rankings persist across independent training seeds and comparable total computational budgets? How do the frozen-reference and deterministic counterparts perform?
-
Can you provide full-dimensional drug-task Wasserstein or sliced-Wasserstein results and explicitly label the coordinate space used by every reported metric?
-
How is constructed and extended to held-out LARRY source cells? What accuracy is obtained using these weights and terminal-class readouts without learned transport?