Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

A compact pretrained model infers ODE vector fields from noisy trajectories in one pass, outpacing a larger symbolic baseline on reconstruction.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 03:11 UTC pith:UM4NGEGV

load-bearing objection FIM-ODE offers a plausible small-model alternative to symbolic amortized ODE inference, but the headline numbers mix denominators and the abstract oversells; it deserves a serious referee, not a desk reject. the 4 major comments →

arxiv 2602.08733 v2 pith:UM4NGEGV submitted 2026-02-09 cs.LG

Foundation Inference Models for Ordinary Differential Equations

Maximilian Mauel , Johannes R. H\"ubers , David Berghaus , Patrick Seifner , Ramses J. Sanchez This is my paper
classification cs.LG
keywords vector field inferenceneural operatorspretrainingsystem identificationordinary differential equationszero-shot learningtrajectory dataamortized inference
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes FIM-ODE, a pretrained foundation model that maps noisy, irregularly sampled trajectory data directly to a local estimate of an ordinary differential equation's vector field in a single forward pass. Its central claim is that a model trained only on low-degree polynomial ODEs can generalize to broader system classes, and that this local neural-operator representation can match or beat a global symbolic-regression baseline on trajectory reconstruction while staying competitive on generalization. The discovery matters because it suggests that for many real-world inference tasks, the cost of system identification can be shifted from per-dataset optimization to a single pretraining phase, and that simple local structure—state, displacement, squared displacement, and time gap—carries enough information to reconstruct dynamics in data-rich regions.

Core claim

The paper's central claim is that vector-field inference for ODEs can be amortised: a model pretrained on sparse random polynomial vector fields (degree at most three, dimensions one to three) learns a map from noisy trajectory context to a continuous local estimate of the underlying field at any query point. On ODEBench trajectory reconstruction, this estimator achieves variance-weighted R-squared greater than 0.9 in 84.4% of systems at zero noise, compared to 63.1% for a global symbolic baseline, and it remains competitive on generalization to new initial conditions. Pretraining also serves as an initialisation that, after brief finetuning, outperforms per-dataset neural and Gaussian-proce

What carries the argument

The load-bearing component is the transition-based input representation: each consecutive observation pair is encoded as the current state, the displacement, the element-wise squared displacement, and the inter-observation time gap. Together these act as a finite-difference estimate of the local vector field plus a second-moment feature capturing noise and local scale. A transformer encoder forms a permutation-invariant context; a functional attention decoder queries this context at arbitrary spatial locations, producing a continuous neural-operator field estimate. An auxiliary head predicts a per-location heteroscedastic log-variance that weights the Laplace likelihood loss, down-weighting

Load-bearing premise

The vector field in the region visited by the data is determined well enough by local transition statistics—state, displacement, squared displacement, and time gap—that a model trained on polynomial fields transfers to non-polynomial systems in that same region.

What would settle it

Take a conservative system (e.g., an undamped pendulum) and a single noisy trajectory; run FIM-ODE and evaluate the inferred vector field on a grid spanning unvisited regions. If the field error grows sharply away from the trajectory and closed orbits are not preserved, the locality premise fails, as the paper already notes for the pendulum.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • In data-rich regimes, a local estimator can beat symbolic regression on trajectory reconstruction even when the system's functional form lies outside the training prior, as shown on a rational reaction-network model.
  • Pretraining on simple polynomial ODEs transfers to non-polynomial systems, including trigonometric and rational vector fields, when the relevant behavior is controlled by local features in the visited region.
  • A brief finetuning phase turns the pretrained model into a strong per-dataset estimator, outperforming modern neural and GP baselines without requiring derivative estimation or solver-based training.
  • Performance on low-data benchmarks is strongly sensitive to the noise realisation, so single-seed evaluations can misrepresent zero-shot capability; the paper reports large variance across 100 seeds.
  • The local representation is not universal: for conservative systems such as the frictionless pendulum, the model fails to preserve closed-orbit geometry and generalises poorly to unseen initial conditions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The local representation suggests an evaluation protocol that reports accuracy only on the data-visited region, since global claims about unvisited state-space are exactly where the model fails; the paper's own pendulum analysis supports this boundary.
  • The transition features resemble a second-order expansion of the flow; a testable extension is whether adding higher-order or memory-dependent features would capture conservative geometry without a richer prior.
  • Because the model can condition on multiple trajectories, it could serve as a fast initialiser for classical system-identification pipelines, proposing a good starting field that a local refinement step then corrects.
  • The authors' proposed stationary-GP prior is a natural next test: it may preserve stability while fixing the sink bias that the polynomial prior induces, and its effect on conservative systems is directly measurable with the pendulum benchmark.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper introduces FIM-ODE, a pretrained transformer-based neural operator for amortized inference of low-dimensional ODE vector fields from noisy, irregularly sampled trajectories. It is pretrained on a synthetic corpus of 600K polynomial ODE systems (degree ≤ 3, dimensions 1–3) and evaluated zero-shot on ODEBench against ODEFormer, on low-data VDP/FHN benchmarks against classical neural/GP baselines, and on CMU MoCap human-motion data. The central claims are that FIM-ODE's local representation matches or improves upon the global symbolic ODEFormer despite a simpler prior and much smaller model/data scale, and that pretraining provides a strong initialization for finetuning. The paper includes candid limitations and extensive appendices, but several reporting choices (excluded systems, selected-noise results, and contradictory table/text statements) undermine the headline claims as currently written.

Significance. If the empirical claims are substantiated, the work is significant: it shows that a relatively small (8M-parameter) local neural-operator estimator, pretrained on a simple polynomial prior, can amortize ODE inference competitively against an 86M-parameter symbolic transformer trained on 50M systems. The architecture is plausible, the transition-feature representation and uncertainty weighting are sensible, and the paper includes useful ablations on dataset size, discretization, and noise-seed sensitivity. However, the central comparison is currently not fully reliable because of a possible denominator mismatch in Table 1, and the paper's own tables contradict several textual claims about finetuning and zero-shot performance. These issues are fixable but require substantial revision and re-analysis.

major comments (4)
  1. [§5.1, Table 1, footnote 2] FIM-ODE is evaluated on 59 ODEBench systems after excluding two systems with dimension > 3, but the ODEformer percentages in Table 1 are not stated to be recomputed on the same 59-system subset. If the ODEformer numbers are taken from d'Ascoli et al. (2024) over all 61 systems, the comparison mixes denominators: 63.1% of 61 is about 38.5 systems, while 84.4% of 59 is about 49.8 systems. Since Table 1 is the main support for the claim of improving upon ODEFormer, the authors must re-evaluate ODEformer on the identical 59-system subset and report per-system results, or clearly state the denominator used for each method.
  2. [§5.2, Table 4] The text claims that finetuning 'yields the best results among all methods considered,' but Table 4 lists GP-DNF with VDP T1/T2 MSE of 0.03/0.04 while FIM-ODE (Finetuned) has 0.16/0.26. The zero-shot sentence 'FIM-ODE consistently outperforms ODEformer in zero-shot mode' is also contradicted by VDP T1 (FIM-ODE 0.89 vs ODEformer 0.22). Moreover, the 'FIM-ODE (Selected noise)' rows are the best of 100 independent trials (Appendix E.2), not representative zero-shot performance; Table 7 reports mean/median VDP T1 MSE of 0.575/0.340. These claims and rows must be corrected, removed, or replaced with distributional statistics.
  3. [Abstract and §5.1] The abstract's claim of 'matching and often improving upon ODEFormer ... across a range of regimes' is not supported by the tables. Table 2 shows no consistent advantage in generalization (e.g., at ρ=0.5, σ=0.03, ODEformer is 32.8% vs FIM-ODE 27.0%), and Table 4 shows zero-shot FIM-ODE is substantially worse than ODEformer on VDP T1. The only strong evidence is Table 1 reconstruction, which is subject to the denominator concern above. The claims should be qualified to the reconstruction task and the regimes where FIM-ODE actually wins, with a discussion of observed failures.
  4. [§5.1, Figure 2, §4.2] The frictionless-pendulum result directly challenges the load-bearing premise of §4.2 that local transition features (state, displacement, squared displacement, time gap) determine the vector field in the data-visited region. For this non-polynomial system, FIM-ODE 'tends to bias the field toward a weak sink near the origin' and 'generalises poorly to unseen initial conditions,' even though it matches the context trajectory. The text nevertheless concludes that 'local estimates can transfer, even when the global functional form is OOD.' This counterexample should be addressed quantitatively, e.g., by reporting ODEBench success rates separately for polynomial and non-polynomial systems, and by tempering the transfer claim.
minor comments (6)
  1. [Appendix B, References] The reference to 'Hoff, J. H. v. t. Studies in Chemical Dynamics' gives the year 1986; the cited work is from 1884. Also 'V olterra' is misspelled in the text and bibliography.
  2. [Appendix A.1] The phrase 'any adhering initial value problem' should be 'any admissible initial value problem' or similar; 'adhering' is not standard terminology.
  3. [Appendix D] The sentence 'After nsteps steps ˆxi' has a redundant word. Also, the description of how initial conditions are synchronized with the observed time grid would benefit from a small diagram or explicit pseudocode.
  4. [Table 6] The text says FIM-ODE shows a 'clearer edge' at the R2 > 0.8 threshold, but at ρ=0.5, σ=0.03 ODEformer achieves 41.0% vs FIM-ODE 40.2%; the description should acknowledge this exception.
  5. [Figure 2 caption] The caption says 'clean context trajectory' although the setting uses σ=0.03 noise and ρ=0.5 subsampling. Please rephrase to 'corrupted context trajectory' or similar.
  6. [General] The notation is sometimes confusing: y is used both for corrupted observations and as a state variable in parts of the appendix. A brief notation table or consistent subscript convention would improve readability.

Circularity Check

0 steps flagged

No circularity: FIM-ODE's central claims rest on external zero-shot benchmark evaluation, not on self-citations or fitted constants.

full rationale

Walking the derivation chain: the pretraining prior p(f) is a distribution over polynomial ODEs with ground-truth vector fields f(x); trajectories are simulated and corrupted; the model is trained by matching predicted vector fields to these ground-truth values at query locations (Section 4.2, loss Lθ). The ODEBench results in Tables 1 and 2 are zero-shot evaluations against an external baseline (ODEformer) and external per-dataset methods, so the headline quantitative claims are not derived from fitted constants or from the paper's own equations. The self-citations to Seifner et al. (2025a) — 'This architecture is adapted from Seifner et al. (2025a)' and 'We therefore follow Seifner et al. (2025a) and use uncertainty weighting' — are architectural provenance, not load-bearing justification; no uniqueness theorem or ansatz is imported from the authors' prior work to force the result. The GP interpretation in Section 4.1 is explicitly descriptive and does not feed into the loss or evaluation. The paper's own limitation statement in Section 5.1 ('FIM-ODE tends to bias the field toward a weak sink near the origin ... fails to preserve the correct conservative geometry') weakens the generalization claim but is an honest empirical failure mode, not a circular step. The possible ODEBench system-set mismatch flagged by the skeptic (59 vs 61 systems) is an evaluation-validity concern, not a circularity of the derivation; no equation or fitted parameter reduces to the prediction by construction. Hence: no significant circularity.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

The paper's contributions are empirical; the ledger shows the many hand-set prior and training choices that shape the claims. The most load-bearing is the local-sufficiency axiom (axiom 3), which the paper itself partially refutes. The 'Selected noise' parameter (free parameter 6) is a selection maximum, not a fitted constant in the circularity sense, but it materially affects the headline low-data numbers.

free parameters (7)
  • Pretraining prior: max polynomial degree p=3 = 3
    Bounds the hypothesis class; chosen by hand to cover canonical ODEs (Lorenz, VDP), affects the OOD degree-6 generalization claim.
  • Coefficient scale and global scaling = cA~N(0,1), s~Uniform[0,2]
    Controls vector-field magnitudes and stability of sampled systems; ad hoc prior design.
  • Trajectory rejection threshold = δ_reject=100
    Divergent systems discarded; biases prior toward bounded trajectories; acknowledged as a limitation.
  • Corruption ranges = σ~Uniform[0,0.06], ρ~Uniform[0,0.5]
    Copied from ODEformer protocol; determines robustness claims.
  • Normalization target time-gap = Δτ_target=0.01
    Centers inter-observation times; arbitrary scale choice affecting all predictions.
  • Low-data zero-shot noise seed = best of 100 trials (selected row)
    The main-table 'Selected noise' row is a maximum over seeds; mean VDP T1 MSE is 0.575 vs 0.02 selected (Table 7).
  • Finetuning hyperparameters = nsteps=25, nIC=⌊2ℓ/25⌋, σ_i=Δt/5, 200-800 epochs
    Task-specific settings for the neural-ODE finetune; not justified by ablations.
axioms (6)
  • standard math Picard–Lindelöf theorem guarantees unique local solutions for locally Lipschitz polynomial vector fields.
    Invoked in §3 to justify simulating and matching trajectories.
  • domain assumption Target observations y are noisy state measurements of an autonomous first-order ODE; multiplicative noise plus Bernoulli subsampling is a faithful corruption model.
    §4.1; the entire benchmark rests on this generative assumption.
  • domain assumption The vector field is locally identifiable from transition features (state, Δy/Δτ, Δy², Δτ) in the visited region, and this local estimate transfers OOD.
    §4.2 'Transition-based input representation'; load-bearing and contradicted by the frictionless-pendulum generalization failure in §5.1/Figure 2.
  • domain assumption Reversible instance normalization (per-dimension zero mean/unit variance) and delta-log time centering preserve enough information to predict vector fields; denormalization by the chain rule is exact.
    Appendix C.2; if the observed trajectory is non-ergodic or chaotic, the normalization constants depend on the window and can distort the field.
  • domain assumption Variance-weighted R² > 0.9 on integrated trajectories is a valid measure of vector-field inference quality.
    §5/Appendix E.1; threshold choice (0.9 vs 0.8) changes relative ranking.
  • domain assumption Neural-ODE finetuning on context trajectories with multiple shooting initial conditions transfers to held-out initial conditions and subjects.
    Appendix D; the MoCap finetuning claim depends on this transfer.

pith-pipeline@v1.3.0-alltime-deepseek · 24925 in / 21738 out tokens · 215845 ms · 2026-08-03T03:11:01.484170+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Foundation Inference Models for Ordinary Differential Equations." pith.science (2026). https://pith.science/paper/UM4NGEGV

@misc{pith2026260208733,
  author       = {Pith},
  title        = {Pith review of: Foundation Inference Models for Ordinary Differential Equations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UM4NGEGV}},
  note         = {Machine review of arXiv:2602.08733}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Ordinary differential equations (ODEs) are central to scientific modelling, but inferring their vector fields from noisy trajectories remains challenging. Current approaches such as symbolic regression, Gaussian process (GP) regression, and Neural ODEs often require complex training pipelines and substantial machine learning expertise, or they depend strongly on system-specific prior knowledge. We propose FIM-ODE, a pretrained Foundation Inference Model that amortises low-dimensional ODE inference by predicting the vector field directly from noisy trajectory data in a single forward pass. We pretrain FIM-ODE on a prior distribution over ODEs with low-degree polynomial vector fields and represent the target field with neural operators. FIM-ODE achieves strong zero-shot performance, matching and often improving upon ODEFormer, a recent pretrained symbolic baseline, across a range of regimes despite using a simpler pretraining prior distribution. Pretraining also provides a strong initialisation for finetuning, enabling fast and stable adaptation that outperforms modern neural and GP baselines without requiring machine learning expertise.

Figures

Figures reproduced from arXiv: 2602.08733 by David Berghaus, Johannes R. H\"ubers, Maximilian Mauel, Patrick Seifner, Ramses J. Sanchez.

Figure 1
Figure 1. Figure 1: Synthetic data generation (left) and FIM-ODE architecture (right). sparse measurements. Let D∗ = {(y ∗ 1 , τ ∗ 1 ), . . . ,(y ∗ L , τ ∗ L )} denote a dataset of L observations recorded at irregular time points 0 ≤ τ ∗ 1 < · · · < τ ∗ L . We assume that each obser￾vation y ∗ i ∈ R d corresponds to a noisy measurement of the (hidden) state x(τi) of a dynamical system governed by the ODE above (Eq. 1), with a… view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of ODEformer and FIM-ODE on two ODEBench systems. Each model infers a vector field from a single corrupted context trajectory, obtained by subsampling the ground truth (ρ = 0.5) and adding noise (σ = 0.03). The clean context trajectory is shown in green in the left column. We then integrate the inferred vector fields from two initial conditions specified by ODEBench to assess reconstruction and … view at source ↗
Figure 3
Figure 3. Figure 3: Magnitude statistics of vector field points as a function of relative distance to bounding box borders for 1D ODEs. The 1D case exhibits unique behavior due to outlier systems (see [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Magnitude statistics of vector field points as a function of relative distance to bounding box borders for 2D ODEs. Dataset Sizes. The pretraining dataset contains 600,000 polynomial ODE systems spanning dimensionality 1, 2, and 3 (80,000 one-dimensional, 210,000 two-dimensional, and 310,000 three-dimensional systems). A validation set is generated following the same distribution at 10% of the training set… view at source ↗
Figure 5
Figure 5. Figure 5: Magnitude statistics of vector field points as a function of relative distance to bounding box borders for 3D ODEs. 0 5 10 15 20 25 30 x -125 -100 -75 -50 -25 0 25 50 75 100 125 f ( x ) f(x) |f(x)| [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Example 1D outlier ODE. Vector field (blue) and magnitude of vector field (orange). The magnitude changes abruptly when the vector field crosses the x-axis, creating the distinctive pattern observed in 1D statistics. C. Architecture and Training Details This appendix provides comprehensive details on the FIM-ODE architecture, normalization schemes, and training proce￾dures. C.1. Architecture Specifications… view at source ↗
Figure 7
Figure 7. Figure 7: Weighted L1 training and validation losses during pretraining. Loss is weighted according to learned uncertainty estimates. Time 0 100 200 300 400 L1 Loss Train L1 Time 50.0 52.5 55.0 57.5 60.0 62.5 65.0 67.5 L1 Loss Validation L1 [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Unweighted L1 training (left) and validation (right) losses during pretraining. Unweighted Loss Comparison [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Uncertainty estimates for training and validation (left), and gradient norm evolution (right) during pretraining. −2 −1 0 1 2 3 4 5 6 Comp 1 −2.0 −1.5 −1.0 −0.5 0.0 0.5 1.0 1.5 2.0 Comp 2 −12 −10 −8 −6 −4 −2 0 2 Comp 3 Ground truth Base Finetuned −2 0 2 4 Comp 0 Ground truth Base model Finetuned model 0 2 Comp 1 Ground truth Base model Finetuned model 0.0 0.5 1.0 1.5 2.0 2.5 3.0 Time −10 −5 0 Comp 2 Ground… view at source ↗
Figure 10
Figure 10. Figure 10: Finetuning task: MoCap 35 with short context length context trajectories as well as separate validation and testing sequences. We choose the best model based on the neural ODE loss for the validation trajectories. For VDP and FHN, the training and validation trajectories are identical. In less data-poor settings than VDP and FHN, such as the MoCap tasks, we find it useful to regularize finetuning by injec… view at source ↗
Figure 11
Figure 11. Figure 11: Best-performing random samples from 100 independent noise/sampling trials for each benchmark. Blue dots indicate training observations, black lines show ground truth, and red dashed lines show model predictions. benefit diminishes with increasing dimension. For 1D systems, the median RMSE decreases by a factor of approximately 4 when given 9 trajectories instead of 1. This improvement factor reduces to 1.… view at source ↗
Figure 12
Figure 12. Figure 12: Comparison of FIM-ODE and ODEformer on polynomial ODEs (degree ≤ 3). Left: reconstruction performance. Right: generalization to new initial conditions. Both evaluated across dimensions d ∈ {1, 2, 3}. 60 65 70 75 80 85 90 95 Trajectories 1 Trajectories 9 Trajectories 12 50 Points 100 Points 200 Points 250 Points Number of Points per Trajectory 0 5 % of Trajectories with R 2 > 0.9 [PITH_FULL_IMAGE:figures/… view at source ↗
Figure 13
Figure 13. Figure 13: Reconstruction performance for different trajectory discretizations (npoints) and numbers of input trajectories (K). Higher is better. differences on uncorrupted inputs. This suggests that even moderate-scale pretraining (50,000–100,000 systems) provides sufficient coverage of polynomial ODE dynamics for zero-shot transfer to real-world systems. For generalization, larger datasets provide more consistent … view at source ↗
Figure 14
Figure 14. Figure 14: Reconstruction performance as percentage of trajectories with R 2 > threshold, shown separately for each dimension. Brighter colors indicate better performance. Results shown for all tested model inputs (combinations of npoints and K). 50 100 200 250 Number of Points per Trajectory 0 10 20 30 40 50 60 RMSE 4.09 3.91 3.89 3.80 2.74 2.63 2.44 2.35 2.34 2.23 2.37 2.23 1 Trajectory 2 Trajectories 3 Trajectori… view at source ↗
Figure 15
Figure 15. Figure 15: RMSE of vector field predictions for different trajectory discretizations and numbers of input trajectories. Lower is better. 50 100 200 250 Number of Points per Trajectory −0.50 −0.25 0.00 0.25 0.50 0.75 1.00 Cosine Similarity 0.48 0.53 0.53 0.53 0.81 0.84 0.82 0.82 0.82 0.85 0.84 0.84 1 Trajectory 2 Trajectories 3 Trajectories [PITH_FULL_IMAGE:figures/full_fig_p023_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Cosine similarity between predicted and true vector fields for different trajectory discretizations and numbers of input trajectories. Higher is better. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: RMSE of vector field predictions per dimension for 200-point trajectory discretization. Lower is better. Dimension 1 Dimension 2 Dimension 3 Dimensionality −0.4 −0.2 0.0 0.2 0.4 0.6 0.8 1.0 Cosine Similarity 1.00 0.63 0.41 1.00 0.90 0.70 1.00 0.92 0.72 1 Trajectory 2 Trajectories 3 Trajectories [PITH_FULL_IMAGE:figures/full_fig_p024_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Cosine similarity between predicted and true vector fields per dimension for 200-point trajectory discretization. Higher is better. 10k 50k 100k 200k 10 20 30 40 50 60 70 80 90 Reconstruction 10k 50k 100k 200k 10 20 30 40 50 60 70 80 90 Generalization 1 Trajectory 9 Trajectories Size of Training Set % of Trajectories with R 2 > 0.9 [PITH_FULL_IMAGE:figures/full_fig_p024_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Reconstruction (left) and generalization (right) performance as a function of training dataset size. Performance measured as percentage of trajectories with R 2 > 0.9 on in-distribution polynomial ODEs. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: ODEBench reconstruction performance for models trained on 10k, 50k, 100k, and 200k polynomial ODEs. σ denotes Gaussian noise and ρ denotes the dropout ratio used for irregular sampling. 0 20 40 60 80 100 % Accuracy (R2 > 0.9) 200k 100k 50k 10k ρ = 0.0 0 20 40 60 80 100 % Accuracy (R2 > 0.9) ρ = 0.5 σ = 0.0 σ = 0.03 σ = 0.05 [PITH_FULL_IMAGE:figures/full_fig_p025_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: ODEBench generalization performance for models trained on 10k, 50k, 100k, and 200k polynomial ODEs. σ denotes Gaussian noise and ρ denotes the dropout ratio used for irregular sampling. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_21.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. HyperODE: Zero-Shot Surrogate for Simulation and Inference of Dynamical Systems

    cs.LG 2026-08 conditional novelty 7.0

    A single 13k-parameter hypergraph surrogate trained on simple mass-conserving ODEs predicts calibrated trajectory quantiles and infers parameters across unseen compartmental families, system sizes, interaction orders,...

Reference graph

Works this paper leans on

3 extracted references · 1 linked inside Pith · cited by 1 Pith paper

  1. [2]

    Delahunt, C

    URL https://openreview.net/forum? id=TzoHLiGVMo. Delahunt, C. B. and Kutz, J. N. A toolkit for data-driven discovery of governing equations in high-noise regimes. IEEE Access, 10:31210–31234, 2022. 9 Foundation Inference Models for ODEs Dooley, S., Khurana, G. S., Mohapatra, C., Naidu, S. V ., and White, C. Forecastpfn: Synthetically-trained zero-shot for...

    arXiv 2022

  2. [2022]

    Newton, I.Philosophiae Naturalis Principia Mathematica

    URL https://openreview.net/forum? id=KSugKcbNf9. Newton, I.Philosophiae Naturalis Principia Mathematica. cf. especially Liber de Motu I Sects. II, III, VII. London, 1687. Pal, A., Ma, Y ., Shah, V ., and Rackauckas, C. V . Opening the blackbox: Accelerating neural differential equations by regularizing internal solver heuristics. InInterna- tional Confere...

    arXiv 2021

  3. [2024]

    Berghaus, D., Seifner, P., Cvejoski, K., Ojeda, C., and S´anchez, R

    URL https://openreview.net/forum? id=f4v7cmm5sC. Berghaus, D., Seifner, P., Cvejoski, K., Ojeda, C., and S´anchez, R. J. In-context learning of temporal point pro- cesses with foundation inference models.arXiv preprint arXiv:2509.24762, 2025. Bernoulli, J.Extrait de la R ´eponse de M. Bernoulli `a M. Herman, dat´ee de Basle le 7. Octobre 1710.M ´emoires d...

    Pith/arXiv arXiv 2025