Geometric Flow Grounding: A Unified Manifold Decoupling Framework for Dynamics Discovery and Verification

Summary

Geometric Flow Grounding (GFG) proposes to learn continuous dynamics that are, by construction, tangent to a learned data manifold. Two mechanisms carry the framework: a differentiable Neural Tangent Projection that uses a matrix-free Jacobian–vector product of the state decoder to force predicted velocities into the column space of $J_G$, and a dual-stream vector-quantized architecture that separately codes static topology and dynamic evolution primitives. The conceptual move — reusing a single tangent-projection operator as (i) a numerical stabilizer for equation discovery from sparse ODE snapshots, (ii) a geometric regularizer for RNA-velocity inference, and (iii) a training-free consistency critic for AI-generated video — is genuinely novel, and the empirical breadth is substantial. Theorem 3.2 gives a clean decomposition that motivates the projection as both an inference denoiser and a training regularizer on the learned tangent geometry, and the JVP realisation has a real efficiency advantage over chart-based (NM-ODE) or full-Jacobian (RFM) alternatives.

The paper's central limitation is that the JVP-plus-VQ package is stretched across three fields whose failure modes are different, and the empirical narrative repeatedly overshoots what the tables actually show. Standalone GFG is dominated by NSG-VD on every video-detection metric; only the fused GFG+NSG-VD wins Recall/Accuracy/AUROC, and even it slightly regresses F1 — yet the text describes this as "best overall". Table 3 is not a clean sweep either, Table 2's coefficients are close but not "exact", the phase-ambiguity story only covers half the sampled $\Delta t$ range, and Table 10's Sora F1 for the fused method is arithmetically impossible on the stated balanced test set. Several definitions (ICCoh naming, motif codebook size, $F_i$ length) contradict themselves across the manuscript. The theory is largely correct but Proposition A.1 identifies the tangent space with the column space of $J_G$ without stating the required rank/immersion condition. More consequentially, when the described pipeline is reimplemented from the manuscript (the promised repository is not released), the Case-1 headline table does not reproduce (MSE ~245x higher than reported, VPT ~74% lower), the Case-2 ablation ordering inverts (removing VQ has essentially no effect, removing JVP is the destructive ablation — the opposite of the paper), the $\lambda>0$ vs $\lambda=0$ gap does not appear, and the time-dependent extension in Table 7 is architecturally incompatible with the described autonomous latent field.

Strengths

• Conceptual contribution. A single geometric operator — orthogonal projection onto the column space of a learned decoder Jacobian, evaluated via a matrix-free JVP — is used to serve three otherwise unrelated goals (equation discovery, RNA velocity, deepfake verification). The unification is clean and the "projection residual as physics critic" framing for video is a nice conceptual move.
• Efficiency of the primitive. The JVP realisation avoids explicit Jacobian instantiation and chart construction; Table 9's controlled comparison against NM-ODE and RFM makes the algorithmic advantage tangible (single matrix-free directional derivative vs. sequential autograd or repeated tangent projections). The core operators (JVP tangent projection, orthogonal projection $P_J$ via $d$ JVP calls, RK2 latent integration) are implementable exactly as described.
• Theory grounded in the intended computation. Theorem 3.2 states the tangent-projection error decomposition with respect to the learned tangent space rather than an unknown ideal manifold, and the appendix's imperfect-manifold discussion (A.2) connects the residual to Wedin/Davis–Kahan-type subspace error, which is the right vocabulary for the claim being made.
• Case-1 sparse structure is real. Even in a reimplementation whose absolute error is much larger than reported, GFG recovers the correct sparse term structure of the circular-motion ODE ($\dot x_1$ contains only $x_2$ with a near-unit negative coefficient; $\dot x_2$ contains $x_1$ with a near-unit positive coefficient), while the four described baselines are known to fail these systems for reasons the paper analyzes correctly.
• Interpretability primitive is set up cleanly. The mean-attention-per-cluster analysis of the velocity codebook, plus decoding of atoms to gene expression for GO enrichment, is the right way to test whether VQ codes carry biological structure — a rare setup that could be verified numerically if the Mouse Brain data and codebook seed were released.
• Robustness under compression. Table 11's ~96% AUROC across $R=2$ and $R=4$ downsampling is an appealing sanity check that the residual is not driven by high-frequency pixel artifacts (though the compression operator needs to be pinned).

Weaknesses

• Video headline is systematically overstated. Standalone GFG (R 73.6 / Acc 88.2 / F1 81.5 / AUROC 93.4) is worse than NSG-VD on every metric in Table 4; the fused GFG+NSG-VD wins Recall/Accuracy/AUROC but loses F1 (88.9 vs 90.9) — yet the text claims "best overall" while simultaneously conceding the F1 loss. The honest headline is that GFG contributes a complementary geometric signal to NSG-VD, not that it establishes a new state of the art.
• Table 10 Sora column contains impossible numbers. For GFG+NSG-VD, Recall = 84.42, Accuracy = 91.16, F1 = 53.65 cannot all be true on a balanced 500/500 test set: with $R = 84.42$ and Accuracy $= 91.16$ (~88 total errors, so at most ~10 FP), F1 is $\approx 90.6$, not 53.65. In the other direction, F1 = 53.65 would require FP $\approx 652$ > 500 available real videos. A parallel inconsistency exists for GFG(O) Sora ($R = 66.92$ with Accuracy $99.72$).
• Equation-recovery claim is overstated. The "exact equation coefficients" assertion is contradicted by Table 2 itself ($\dot x_1 = -0.9737\,x_2$, $\dot x_2 = 1.0188\,x_1$ vs ground truth $\pm 1.0$), and in a reimplementation the recovered coefficients drift further ($-0.997$ and $+0.899$). The correct-sign correct-sparsity claim is defensible; "exact" is not.
• Phase-ambiguity explanation covers only part of the sampled data. With $\omega = 1$ and $\Delta\theta \sim U(120^\circ, 240^\circ)$, only the upper half of the increments yield rotations exceeding $\pi$, so the ambiguity argument does not apply to the smaller gaps yet the baselines fail on those too.
• RNA narrative is inconsistent with Table 3. "Consistently outperforms" and "significant Recall gains" both overstate: Case 2 defines no Recall metric, and on Dentate Gyrus veloVI's VeloCoh (0.621) is bolded above GFG (0.473 ICCoh). A hedged "best on most cells, particularly on CBDir" would match the numbers.
• Ablation/sensitivity use an unlisted dataset. Case 2 declares four benchmarks (MouseBrain, Dentate Gyrus, Retina, Erythroid) but Tables 5 and 6 evaluate on Pancreas (and Mouse Brain). Pancreas's role vs the four headline benchmarks is unexplained; this risks reading as a favorable-surface choice, and — as the reproducibility section notes — the conclusions do not survive a faithful reimplementation on that dataset.
• "Eliminates integration error" is scope-broader than the algorithm supports. JVP projection guarantees zero instantaneous normal component; a discrete Euler/RK step in ambient space along a tangent vector still departs a curved manifold. What actually eliminates drift is that Case 1 integrates in the latent space and decodes via $G$, and that scope is not attached to the strong "eliminates" statement in the contributions.
• Theory has a missing rank condition. Proposition A.1 identifies $T_{G(z)}M$ with $\mathrm{Range}(J_G(z))$ without invoking the immersion/full-column-rank condition that this identification requires. Assumption 3.1 provides it but the proof does not carry it through.
• Video pipeline definitions contradict. For $L = 10$ input frames, the paper defines the final feature $F_i$ as the pixel-wise difference of first-order variations (which yields 9 maps) but then states $T = 8$ (the acceleration length). Either $F_i$ is $\Delta \delta$ or $\delta_{\mathrm{pred}} - \delta_{\mathrm{obs}}$; not both.
• Domain shift on the real side of Case 3. Training reals come from Kinetics-400; validation and test reals come from MSR-VTT. Because the score is a residual against the learned manifold, a real-side domain shift can inflate residuals independently of synthesis artifacts. A source-domain real control (held-out Kinetics-400 reals) would separate the two.
• Recurring terminology and framing slips. ICCoh is expanded as both "Intraclass Correlation" (Sec. 4.2) and "Inter-Cluster Coherence" (Sec. B.2.2) even though the body definition is intra-cluster; the "high-dimensional latent space" phrasing in the introduction inverts the ambient/latent convention set in Preliminaries; "orthogonally represent" is stated with no inner-product measurement; the glial GO terms are said to "align precisely" with "myelination" although neither reported term encodes myelination. Individually minor, but they compound.
• Ridge-regularized fit is not the described orthogonal projection. The kinetic residual is presented as a projection onto $\mathrm{Range}(M)$ but implemented as a ridge fit with $\lambda_{\mathrm{reg}} = 10^{-2}$; the residual is therefore shrunken relative to the true orthogonal-projection residual (small bias, but the identification should be qualified).

Reproducibility & code

• Headline Case-1 numbers do not reproduce from the description. A from-scratch implementation of the described GFG pipeline yields one-step MSE 0.0737 and VPT 6.77 for the 2D circular-motion benchmark, versus the paper's 0.0003 and 26.1 (~245x higher MSE, ~74% lower VPT). Recovered frequencies are 0.997 and 0.899, and the rollout accumulates visible phase drift after $t \approx 7$ — the very failure mode the paper attributes only to IRK-SINDy. The 8 non-GFG cells of Table 2 could not be reproduced because none of the four described baselines is a standard shared implementation with a canonical hyperparameter set; the paper does not pin one.
• Ablation ordering (Table 5) inverts under a faithful reimplementation. On Pancreas, reproducing 'w/o JVP', 'w/o VQ', and Full yields CBDir 0.005 / 0.192 / 0.202 (5-run means) versus the paper's 0.405 / 0.125 / 0.588. The full-model value is ~3x lower than reported, and the ordering is reversed: 'w/o VQ' has essentially no effect (within one std of Full), while 'w/o JVP' is destructive. Because the swap definitions are given only in prose ("w/o JVP" — replaced with what? "w/o VQ" — a continuous head? a single vector? codebook removed?), the source of the divergence cannot be pinned down.
• Sensitivity effect (Table 6b) is absent in reimplementation. CBDir at $\lambda=0$ is 0.105 while $\lambda=1$ is $-0.058$ and $\lambda=50$ is $-0.016$; the coherence columns saturate at ~1.0 for every setting. The paper's central claim that turning on the geometric weight produces a "significant performance leap" is not reproducible from the description.
• Time-dependent extension (Table 7) is architecturally incompatible with the described latent field. A reimplementation with the specified time-augmented library produces MSE = NaN and a discovered ODE with no time-dependent terms. The reason is structural: the described latent velocity predictor $V(z_s) = E_v(D_s(z_s))$ is autonomous by construction, so "positional embeddings of time" have nowhere to enter. The manuscript never operationalizes this conditioning.
• Coherence metrics do not discriminate. In the reimplementation, VeloCoh and ICCoh sit in 0.99–1.00 across every configuration (Full, w/o JVP, w/o VQ, every codebook size, every $\lambda$), so near-perfect values on these columns of Table 3 or Table 5 are not evidence of a good velocity field. The paper does not surface this saturation.
• Interpretability on the (external) Mouse Brain dataset is unavailable. Mouse Brain and Retina must be downloaded from external sources not pointed to by the release; on the substituted Pancreas dataset, atom activation is nearly uniform (max single-cluster share 0.202 vs 0.125 uniform), so the Atom 4 / Atom 11 lineage-disentanglement contrast could not be reproduced. Combined with the unpinned Case-2 seed, the specific atom indices in Sec. 4.2 and Sec. D.2.1 are seed-dependent and not reproducible from the description.
• Codebook initialization is unspecified and silently produces total collapse. With the default uniform(-1/K, 1/K) VQ init, Case 1 collapses on 5/5 seeds (reconstruction MSE pinned at the data variance, VQ loss $=0.0$, MSE = NaN); data-dependent initialization is required for the model to train at all. Because the paper does not specify the init, a reproducer following it can silently land on nonsense.
• Case 3 is data-blocked in a paper-only reconstruction. The GenVideo corpus and NSG-VD are external; no download script or NSG-VD interaction code accompanies the description, so Tables 4/10/11 could not be reproduced. Details that would make them checkable are also unstated: the SVD decoder's training-time status (frozen or fine-tuned), the exact score construction ($-\epsilon_\theta$ vs $-\epsilon_\theta/\sigma_t$; whether the UNet input is $X_i$ or noised $X_i$; $\epsilon$- vs $v$-prediction), the Eq. 55 MMD fusion arithmetic, and the compression operator in Table 11.
• Motif codebook size is contradictory across sections. Sec. C.1 says $L = 16$ "as determined by sensitivity analysis", but Table 6(a) peaks at $L = 32$ and Sec. C.2.2 pins codebook_size = 32. A reader cannot tell which value produced Tables 2/3/4.
• Baseline configurations are cited but not pinned for Cases 1 (SINDy/IRK-SINDy/MNN/S-MNN), 2 (scVelo/veloVI/DeepVelo/TIvelo/Graphvelo), and 3 (STIL/NPR/TALL/DeMamba/NSG-VD); it is unclear whether the numbers are re-run in this codebase or transcribed from prior papers.
• Appendix D lacks the setup needed to reproduce it. The time-varying ODE (Table 7), 3D Hamiltonian (Table 8), Burgers (Sec. D.1), and scalability comparison (Table 9) each omit critical details — time-embedding form, initial conditions, augmented SINDy library, PDE viscosity/BCs/integrator/library, and hardware / warm-up / timing protocol. Peripheral pipeline details (the ridge closed form, the GO enrichment tool and gene universe, the compression operator, the Case-3 fusion arithmetic) are likewise omitted.
• The promised repository is not shipped with the submission. Without it — and with the reproducibility gaps above — the paper's main empirical results cannot presently be verified.

Recommended Changes

Essential

• Release a runnable code repository covering all three cases end-to-end, including seeds for Case 2, VQ codebook initialization, the exact "w/o JVP" and "w/o VQ" swap definitions used for Table 5, the SVD decoder's training-time status, the Eq. 55 fusion arithmetic, the NSG-VD interaction code, and data-download pointers for Mouse Brain, Retina, GenVideo, Kinetics-400, and MSR-VTT.
• Investigate the Case-1 headline discrepancy. Either release the exact training script and seed that produced Table 2's MSE = 0.0003 and VPT = 26.1, or reconcile the paper's numbers with the ~245x MSE / ~74% VPT gap observed in a from-scratch reimplementation. Report per-seed MSE and VPT so a reproducer can distinguish a fragile optimum from a robust one.
• Reconcile the Case-2 ablation and sensitivity claims. Publish per-seed CBDir for Full, w/o JVP, w/o VQ, and every entry of Table 6, so the claimed w/o-VQ > w/o-JVP ordering and the $\lambda=0$ vs $\lambda>0$ gap can be checked against the current numbers, which invert both.
• Fix the time-dependent extension by specifying exactly where and how time enters the latent velocity predictor $V(z_s)$; the current description gives $V$ no time input, and a reimplementation returns an autonomous field with no time-dependent SINDy terms.
• Correct the Sora column in Table 10 for GFG+NSG-VD and GFG(O), and verify the rest of Table 10 with the arithmetic check (Recall × Precision → F1 vs Accuracy on a balanced 500/500 test set).
• Re-frame the Case-3 headline to acknowledge that standalone GFG is dominated by NSG-VD on every Table-4 metric and that even the fused system regresses F1; keep the "complementary geometric signal" framing already present in the same paragraph.
• Reconcile the motif codebook size by stating a single value of $L$ used for Tables 2/3/4 (or tabulating per-case values) and aligning Sec. C.1's rationale with the codebook-size peak in Table 6(a).
• Fix or narrow the phase-ambiguity explanation in Sec. 4.1 to $\Delta t > \pi/\omega$ steps, or supply an alternative mechanism for the smaller-gap failures.
• State the immersion / full-rank condition in Proposition A.1 so that $T_{G(z)}M = \mathrm{Range}(J_G(z))$ is a legitimate identification.
• Pin and document the VQ codebook initialization, and note the collapse mode observed with the default uniform init.
• Resolve the video feature-length inconsistency by rewriting the $F_i$ definition so it matches $T = 8$: state whether $F_i$ is the residual of accelerations or of first-order variations, and align $T$ accordingly.
• Justify the ablation dataset choice. Either add ablation and sensitivity tables on one or more of the four headline benchmarks (MouseBrain, Dentate Gyrus, Retina, Erythroid Lineage) or explain why Pancreas is the right surface.

Suggested

• Qualify the "eliminates integration error" claim to specify that drift removal follows from latent-space integration + decoding via $G$, not from tangent projection alone at each ambient step.
• Tone down "consistently outperforms" in the Case-2 narrative to reflect that GFG wins on most, not all, cells of Table 3 (losing VeloCoh on Dentate Gyrus), and remove the "significant Recall gains" phrase — Case 2 has no Recall metric.
• Surface the VeloCoh/ICCoh saturation by reporting these metrics for every ablation/sensitivity configuration alongside CBDir, so readers can see that they saturate near 1.0 and that CBDir is the discriminating column.
• Report a source-domain real control for Case 3 (residual distribution on held-out Kinetics-400 reals) to rule out the Kinetics-400 → MSR-VTT domain shift as a confound.
• Standardize ICCoh naming ("Intra-Cluster Coherence") across Sec. 4.2 and Sec. B.2.2, and fix the "high-dimensional latent space" phrase in the introduction to "high-dimensional ambient space".
• Restate the "orthogonally represent" claim as "disentangle" unless a numerical inner-product measurement between the neuronal and glial primitives is added.
• Re-attribute the myelination interpretation to the oligodendrocyte driver genes rather than the two reported GO terms.
• Add a one-line note that $J_{\mathrm{Decoder}} \cdot v_{\mathrm{score}}$ in Sec. 4.3 is realised via JVP so the video pipeline's matrix-free framing is not superficially contradicted by the equation.
• Specify Case-3 SVD training-time status, the compression operator in Table 11, the GO tooling / gene universe / correction, the ridge closed form, and the Eq. 55 fusion arithmetic so peripheral details do not each require reverse-engineering.
• Report hardware and timing protocol for Table 9 so the millisecond-scale runtime comparison against NM-ODE and RFM is reproducible.
• Soften "insensitive to hyperparameter tuning" to "stable across $\lambda \in [1, 10]$", reflecting the $\lambda = 0$ collapse the paper itself reports.
• Pin baseline configurations (versions, hyperparameters, and whether numbers are re-run or transcribed) for all three cases in a per-baseline table.