-
The code in Appendix C suggests the reaction field cannot affect the reported results.
In forward, d_x = v: spatial trajectories depend only on the velocity network, never on g.
sample() returns z1[..., :dim], discarding the log-weights. All sample-based metrics (EMD, void leakage, kinetic action, NFE of the x-trajectories) are therefore computed on unweighted samples from a plain velocity flow.
For both toy datasets, getitem sets mass_ratio = 1.0, so m_t = exp(t·10⁻¹⁰) ≈ 1. The velocity loss is then ordinary CFM MSE.
If no log_density_fn is set, g_t = 0 and the scalar field is trained toward zero (the same “dormancy” the paper blames on VGFM).
Net effect: on the toy benchmarks, WFR-FM’s velocity learning appears essentially identical to OT-CFM. Consistently, Table 1 shows WFR-FM ≈ OT-CFM/Struct-Coupling in kinetic action (0.2357 vs 0.2384) and EMD (0.3498 vs 0.3431). Yet the NFE differs by roughly 500×.
-
Confound and likely baseline artifact. If the same v-regression gives 44 NFE in one model and 25,018 in another, something other than “decoupling” explains it (undertraining, different augmented state, e.g., log-det channels with tight tolerance for FFJORD-style baselines, different time embedding, a different solver call). Tens of thousands of NFE for dopri5 on a 2D, 8-mode problem is far above what is normally seen (typically tens to hundreds). VGFM, which the authors say has a dormant growth field, is also at 26–45 NFE with similar quality, which suggests low NFE is the normal behavior and the baselines are the anomaly.
-
No stiffness is ever measured. There are no Jacobian eigenvalues, divergence along trajectories, step-size traces, or rejected-step counts. The central causal claim rests on NFE alone.
-
Missing the decisive ablation. WFR-FM with g ≡ 0 (and m_t = 1), i.e., plain CFM with the same backbone, same training budget, same solver, is absent. Also missing: NFE at matched quality, fixed-step Euler/midpoint results (how FM is actually used), and a topology-matched prior such as a K-component Gaussian mixture prior, which sidesteps the obstruction trivially.
-
Headline numbers are selectively presented and partly inconsistent.
“236×” equals 10,381/44, the ratio to the cheapest R = 2 baseline. Against 2-Rectified Flow, it would be 740× (32,581/44). The intro’s “34×–236×” mixes datasets (670/20 = 34× is checkerboard).
The “97× lower kinetic action” compares against FFJORD/OT-Flow (0.078), which aren’t among the four named baselines. OT-CFM (0.0009) and Div-Aligned (0.0000) are as low as or lower than WFR-FM (0.0008).
On R = 2, WFR-FM’s Sinkhorn EMD (0.3498) is about 1.9× worse than FFJORD/OT-Flow/2-RF (0.19–0.20). On the checkerboard, it has the same void leakage as everyone (≈ 0.63), so all methods apparently fail there (leakage 0.60–0.78), contradicting “robust topological support alignment”.
At R = 5, kinetic action is worse than OT-CFM/Struct/VGFM/Div-Aligned, and EMD differences are within one SD (n = 5).
NFE SDs of exactly 0.0 (WFR-FM 44.0 ± 0.0, 20.0 ± 0.0; VGFM 26.0 ± 0.0) across five independently trained networks with an adaptive solver are surprising and need explanation.
-
Table 1 vs. Table 4 contradict each other. Table 4 marks hyphens as “catastrophic solver failure (timeout/NaN)”, but Table 1 reports finite means for the same cells: OT-CFM R = 5 (27,366), Struct-Coupling R = 2 (20,932), Div-Aligned R = 5 (20,888), and finite leakage for many “failed” cells. Table 4 also says “>30,000” for 2-RF, whereas Table 1 says 30,802.
-
Metrics are undefined. “Void leakage”, “kinetic action” (does it include the g term?), the Sinkhorn EMD configuration, and the checkerboard “void” are never defined in the main text. For a perfect model, leakage should be ≈ 0, so values near 0.6 for all methods are red flags about the metric or the models.
-
D = 512 experiment is under-specified. The data distribution isn’t given, and the “KL-Ablation” is not defined. NaN at D ≥ 64 from log-ratios is surprising in float32 (log-densities are only hundreds), and WFR-FM itself exponentiates the ratio. Gradient variance is not a PŁ test. The text says “no gradient clipping” (Table 3, Fig. 3), but train_single calls clip_grad_norm_(…, train_config.grad_clip).
-
Baseline fidelity and fairness.
“Divergence-Aligned FM (Fan et al., 2025)” cites a NeurIPS paper titled Adaptive divergence regularized policy optimization for fine-tuning generative models, an RL fine-tuning method, which does not obviously match the described “penalize velocity divergence” baseline (to verify).
2-Rectified Flow is a two-stage method, which conflicts with the paper’s “single-stage only” scope.
WFR-FM uses two networks (about 2× parameters, and both are evaluated at every NFE), versus one for baselines. Text says 4 layers; code defaults to 2 residual blocks. No per-method hyperparameter tuning; training is only 100 epochs on 10k points.
“Optimal WFR coupling / Minibatch Unbalanced OT”: the code uses balanced entropic Sinkhorn with argmax rounding, off by default (coupling_mode="independent").
-
Section 5.1 sensitivity claims don’t match data. “Uniform box prior reduces leakage to 0.31 on R = 5” is higher than the 0.0250 in Table 1. The γ/τ grid shows only “reaction loss ~10⁻³” (consistent with a field trained toward zero), and τ is never defined. NFE for the grid is not shown.