Pith. sign in

REVIEW 4 major objections 5 minor 27 references

Curved paths reveal when a flow model is guessing.

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

2026-08-02 23:38 UTC pith:KOLB7YFG

load-bearing objection Conditional OT-FM can be made to flag off-manifold inputs by contrastively curving the vector field; the empirical case is strong, and the far-OOD transfer is real but not structurally guaranteed. the 4 major comments →

arxiv 2602.13061 v2 pith:KOLB7YFG submitted 2026-02-13 cs.LG cs.AIstat.ML

Native Extrapolation Awareness in Flow-Based Conditional Generation

Constantinos Tsakonas , Serena Ivaldi , Jean-Baptiste Mouret This is my paper
classification cs.LG cs.AIstat.ML
keywords Flow Matchingout-of-distribution detectionextrapolation detectionconditional generationoptimal transportcontrastive learningadversarial perturbationsconformal prediction
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.

Flow Matching models are a standard choice for conditional generation, but they silently hallucinate plausible outputs when given conditions that fall off the data manifold, and likelihood scores do not catch this. The paper claims this failure can be designed out of the model itself: train the flow with two contrastive penalties—one on transport-energy magnitude, one on direction—relative to the straight optimal-transport path, using adversarial perturbations as hard invalid inputs. Valid conditions keep nearly straight trajectories, while off-manifold conditions are forced into curved, inefficient paths. The area between the generated path and its chord (the DOT score) then separates valid from invalid inputs, reaching AUROC above 0.98 on synthetic and weather benchmarks and 0.86–0.955 on semantic shifts, at no loss in predictive fidelity and with zero extra inference cost. A reader should care because a model that can flag its own extrapolations is a prerequisite for safe deployment in forecasting, robotics, and clinical settings.

Core claim

The central claim is that extrapolation detection in conditional Flow Matching is not a separate classifier problem but a geometric constraint on the learned vector field. Standard OT-FM minimizes transport energy everywhere, so off-manifold conditions inherit smooth, straight, low-energy paths that look exactly like valid predictions. Diverging Flows adds two margin-based contrastive losses to the OT-FM objective: repulsion pulls the off-manifold velocity magnitude away from the optimal transport geodesic, and curvature pushes its direction orthogonal to that geodesic. With PGD-mined hard negatives inside an ε-ball around each valid condition, the vector field develops a sharp discontinuity

What carries the argument

The moving parts are the learned conditional vector field vθ_t, the optimal-transport geodesic u^OT_t used as an anchor, and two contrastive regularizers: L_repel (an L2 margin on transport-energy deviation) and L_curve (a cosine margin on direction). Hard negatives are generated by projected gradient ascent inside an ε-ball around each valid condition, maximizing the flow-matching loss. At inference the DOT score measures the L1 area between the generated trajectory and the straight chord from x0 to x̂1; this quantity is the detector. The loss combination is what makes the vector field Lipschitz-smooth inside the manifold and deliberately non-smooth at its boundary.

Load-bearing premise

The method assumes that adversarial perturbations inside a small ball around each valid condition are a sufficient stand-in for every off-manifold input a deployed model will meet, so the sharp boundary learned at that ball transfers to real, possibly distant out-of-distribution queries.

What would settle it

Train DiFlo on a source manifold, then evaluate DOT scores on natural off-manifold inputs that are far outside the PGD ε-ball used in training—for example, a condition class with different low-level statistics than anything near the training support, or sensor faults of large magnitude. If AUROC falls toward 0.5 as those inputs move farther from the manifold, the learned divergence is confined to the ε-neighborhood and does not constitute implicit manifold learning.

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

If this is right

  • A single Flow Matching model can act as both a high-fidelity conditional generator and its own extrapolation detector, eliminating auxiliary classifiers or post-hoc pipelines.
  • Prediction quality is not traded away: on ERA5 weather forecasting, DiFlo improves MSE (0.0034 vs 0.0038) and keeps SSIM/PSNR comparable; on MNIST-to-SVHN, FID is essentially unchanged (4.104 vs 4.102).
  • Split conformal prediction turns the DOT score into a decision rule with a finite-sample coverage guarantee; at α=0.05 the weather detector reports 5.20% FPR.
  • PGD-mined negatives generalize across domains where fixed heuristics fail (e.g., rotations work on digits but fail on weather), making the method modality-agnostic.
  • Because detection is read off the same trajectory used for prediction, there is zero additional inference latency, suiting real-time safety-critical use.

Where Pith is reading between the lines

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

  • The load-bearing transfer is from ε-ball adversarial negatives to real deployment-time OOD inputs; the paper's own numbers (KMNIST AUROC 0.860 vs FMNIST 0.955) suggest that boundary sharpness weakens as semantic similarity to the training manifold grows, so one testable extension is measuring AUROC as a function of distance from the manifold.
  • If this geometric mechanism holds, path-efficiency scores could become a general alternative to likelihood-based OOD detection across conditional generative models, including diffusion models with curved schedules, by comparing realized path cost to an optimal-transport baseline.
  • The contrastive regularizer explicitly raises the local Lipschitz constant at the manifold boundary, which could interact with adversarial robustness of the predictor; this side effect is untested and worth probing.
  • The paper's proposed extension to Riemannian optimal transport suggests that the same contrastive losses could detect invalid inputs in non-Euclidean state spaces (rotations, poses, manifold-valued data), where the definition of 'straight' comes from the metric.

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 / 5 minor

Summary. The paper introduces Diverging Flows (DiFlo), a modification of optimal-transport Flow Matching that adds two contrastive losses — L_repel (velocity magnitude) and L_curve (velocity direction) — with hard negatives generated by PGD within an ε-ball, so that off-manifold conditions produce non-straight trajectories. At inference, a divergence-from-optimal-trajectory (DOT) score measures trajectory curvature, and split conformal prediction provides a decision threshold. Experiments on synthetic spiral dynamics, ERA5 temperature forecasting, and MNIST→SVHN style transfer with FMNIST/KMNIST OOD report AUROC 0.981, 0.980, and 0.955/0.860 respectively, while claiming preserved fidelity. The central claim is that a single FM model can both generate and natively detect extrapolation.

Significance. The core idea — actively enforcing inefficient transport for invalid conditions — is novel and well-motivated for safety-critical conditional generation. If the claims hold, the method would remove the need for external OOD classifiers or likelihood heuristics. Strengths include the ablation study (Table 1) isolating L_repel, L_curve, and PGD negative mining; detection evaluated on held-out OOD sets (FMNIST, KMNIST, ERA5 hotspots) not used for fitting; and conformal calibration providing finite-sample coverage guarantees. However, the theoretical motivation is heuristic and the empirical support has gaps: the PGD radius ε is never reported, no seed variances are given, and no experiments test OOD queries beyond the ε-ball. These gaps directly affect the load-bearing claim that DiFlo 'implicitly learns the manifold of valid conditions.'

major comments (4)
  1. [§3.5, Eq. (10)–(11); Algorithm 1; Tables 5–7] The PGD radius ε is a core hyperparameter but is never reported in any experiment. Algorithm 1 does not list ε among inputs, and Tables 5–7 omit it. This is a reproducibility blocker and scientifically important: the contrastive loss is only applied inside B_ε(c), so ε controls the distance of the trained decision boundary from the manifold. The paper should report ε for every task and, more importantly, evaluate detection on OOD queries at multiple distances (including far outside B_ε), since the central safety claim concerns arbitrary off-manifold inputs. Current evidence — KMNIST AUROC 0.860 vs FMNIST 0.955 (Table 3) and Domain Transforms outperforming PGD on both digit tasks (Table 4) — suggests the ε-ball boundary transfers only partially. Without far-OOD experiments or a theoretical bound, the safety claim is not established.
  2. [§3.3, Eq. (5)–(6)] The Lipschitz argument is asserted rather than derived. The statement that spectral bias makes the trained model 'inherently minimize' the Lipschitz constant K is a heuristic; Eq. (6) bounds pointwise velocities, not the transport energy E(v;c) over trajectories, and the text itself acknowledges this. Yet later sections refer to a 'structural' geometric phase transition. Please either supply a rigorous bound connecting local ε-ball penalties to behavior outside the ball, or explicitly reframe §3.3 as intuition rather than proof. As written, the theoretical motivation does not support the strong claim of native, universal extrapolation detection.
  3. [Tables 1–4] All AUROC/FPR/fidelity numbers are point estimates without seed variances. Given the fidelity margins are small (e.g., FID 4.104 vs 4.102, LPIPS 0.2202 vs 0.2172, MSE 0.0034 vs 0.0038), the claim 'without compromising predictive fidelity' is not statistically supported. Report mean ± std over at least 3–5 seeds, with significance tests or confidence intervals, especially for the weather and style-transfer comparisons where the differences are within typical run-to-run noise.
  4. [Appendix B.4, Table 8] DiffPath-6D achieves higher AUROC than DiFlo-PGD on all three detection benchmarks (0.989 vs 0.980 for weather; 0.991 vs 0.955 for FMNIST; 0.929 vs 0.860 for KMNIST). The main text describes this as 'comparable.' Please discuss frankly whether the single-model/zero-overhead property is the only advantage over a dedicated external detector, and whether the 'unified off-manifold detection performance' framing should be qualified. This is particularly relevant because the paper's main benchmark excludes post-hoc methods, so Table 8 is the only direct comparison.
minor comments (5)
  1. [Figure 3 caption vs Table 2] The Figure 3 caption reports MSE 0.0004 and SSIM 0.961, while Table 2 reports MSE 0.0034 and SSIM 0.951 for DiFlo on ERA5. These numbers must be reconciled.
  2. [Algorithm 1] Algorithm 1 input list includes λ, β, η, K but omits ε and the margins m_r, m_c. Add all hyperparameters to the input list for completeness.
  3. [§3.7] The notation 'S_DOT(x(c_test))' is confusing; the DOT score is computed from a generated trajectory, not directly from the condition. Clarify that the score is a function of the ODE solution obtained with condition c_test.
  4. [Eq. (14)] Equation (14) has formatting issues (the trace and Jacobian term are garbled) and the dimension of the random matrix Z is unclear. Please restate the instantaneous change-of-variables formula precisely.
  5. [Table 4] The table header 'Temperature Forecasting | Style Transfer' with 'Hotspots' as a subcolumn is ambiguous. Clarify that 'Hotspots' is the ERA5 anomaly set and that 'Style Transfer' detection is vs FMNIST and KMNIST respectively.

Circularity Check

0 steps flagged

No significant circularity: the detection signal is trained in, but every headline AUROC is measured on held-out OOD inputs not used as training negatives, so the evaluation is not circular.

full rationale

The only plausible circularity candidate is that the DOT score (Eq. 12) measures trajectory curvature, which is exactly the geometric property that the contrastive losses L_repel/L_curve (Eqs. 7-9) are designed to create for off-manifold inputs. Taken alone, that would make detection on the training negatives a restatement of the objective. But the paper never reports detection on the PGD negatives used in training. All headline numbers are on disjoint held-out OOD sets: the synthetic OOD is defined with a buffer (epsilon=0.025) outside the PGD epsilon-balls, FMNIST/KMNIST are never seen during training, and the ERA5 hotspots are a different perturbation modality. The FM-DOT baseline achieving AUROC ~0.5 is an important control: the same score on an unregularized FM has no separating power, demonstrating that the metric is not definitionally an extrapolation detector. The conformal threshold is calibrated only on in-distribution held-out samples, so the FPR numbers are not fit to the OOD data. Self-citations (Rouxel et al. 2024, 2025) appear only in contextual lists of flow-matching applications and are not load-bearing for the central claim. No quoted step reduces Eq. X to Eq. Y by construction; the KMNIST AUROC drop (0.860) is a substantive empirical limitation about generalization, not a circularity.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 1 invented entities

The central claim rests on standard OT-FM theory and on three unproven premises: (i) neural smoothness (spectral bias) creates the silent extrapolation hazard; (ii) PGD negatives in the ε-ball transfer to real OOD inputs; (iii) instantaneous contrastive penalties translate into detectable trajectory curvature. The latter two are supported only empirically and are the part of the paper most likely to fail outside the tested benchmarks. The invented quantity (DOT score) is defined by the paper and validated on held-out data.

free parameters (7)
  • λ (repulsion loss weight) = 0.1 (synthetic), 0.7 (weather), 0.1 (style transfer)
    Chosen by hand/validation per experiment; not swept or theoretically motivated. Directly controls how strongly off-manifold velocities are pushed away in magnitude.
  • β (curvature loss weight) = 0.1 (synthetic CG), 0.05 (synthetic regression), 0.9 (weather), 0.7 (style)
    Controls directional (cosine) divergence of off-manifold velocities; tuned per task.
  • m_r (repulsion margin) = 1.0 (synthetic), 100.0 (weather), 10.0 (style)
    Hinge margin for L2 separation; values differ by two orders of magnitude across tasks, indicating task-specific tuning.
  • m_c (curvature margin) = 0.9 (synthetic/weather), 1.3 (style)
    Cosine-distance margin; tuned per task.
  • ε (PGD perturbation radius) = 0.025 (synthetic); not reported for weather/style
    Defines the ball for hard-negative mining (Eq. 10-11) and, in the synthetic benchmark, the train/test OOD boundary. Missing for two of three real-world experiments.
  • η (PGD step size) = 0.1 (synthetic, weather), 0.2 (style)
    Adversarial update step; chosen per task.
  • K (PGD iterations) = 3 (synthetic), 5 (weather, style)
    Number of adversarial refinement steps; more steps cost more training compute.
axioms (6)
  • standard math OT-FM conditional training minimizes the transport-energy excess E(v;c) = E||v_t(x_t,c) − u_OT||^2 with u_OT = x1 − x0 (straight-line conditional paths).
    Invoked in §3.2-3.3 (Eq. 4) and attributed to Lipman et al. 2022 and Benamou-Brenier; foundational for the claim that standard FM encourages straight trajectories.
  • domain assumption The learned vector field is Lipschitz in the conditioning variable c with small constant K due to spectral bias of neural networks (Eq. 5), so off-manifold conditions near the support inherit near-optimal transport and produce silent extrapolations.
    §3.3 Eq. 5-6. Lipschitz-in-c is asserted for ODE well-posedness (the actual requirement is Lipschitz in x), and the smallness of K relies on Rahaman et al. 2019; both are empirical/heuristic, not proven.
  • domain assumption Valid conditions lie on a lower-dimensional manifold MID; test-time off-manifold conditions are the complement in ambient space.
    §3.3 and §4.1 define the problem this way; real OOD (sensor glitches, semantic shifts) is assumed to be off this manifold.
  • ad hoc to paper PGD negatives within an ε-ball around valid conditions are a sufficient surrogate for deployment-time off-manifold inputs, and the induced vector-field discontinuity transfers to unseen OOD queries.
    §3.5-3.6. The entire detection claim depends on this transfer; the paper's own results show imperfect transfer (KMNIST AUROC 0.860 vs FMNIST 0.955).
  • standard math Calibration and test on-manifold scores are exchangeable, so split conformal prediction yields marginal coverage 1−α.
    §3.7; standard conformal assumption, stated explicitly in the paper.
  • ad hoc to paper The contrastive losses on instantaneous velocities (Eq. 7-9) translate into nonzero trajectory curvature (DOT) for off-manifold conditions at inference, rather than merely shifting the straight path to a different endpoint.
    §3.4-3.6, Eq. 12. No proof is given that the integrated ODE path curves; the paper provides empirical separation instead.
invented entities (1)
  • DOT score (S_DOT): divergence-from-optimal-trajectory independent evidence
    purpose: Inference-time anomaly score that measures the L1 deviation of the generated ODE trajectory from the straight chord between its start and generated endpoint; combined with a conformal threshold it decides whether a conditioning input is on- or off-manifold.
    Evaluated on held-out OOD sets (FMNIST, KMNIST, ERA5 hotspots) that did not contribute to training or calibration; these provide a falsifiable handle in the reported benchmarks. Generality beyond these benchmarks is not established.

pith-pipeline@v1.3.0-alltime-deepseek · 21826 in / 30429 out tokens · 261841 ms · 2026-08-02T23:38:12.871094+00:00 · methodology

0 comments
read the original abstract

The ability of Flow Matching (FM) to model complex conditional distributions has established it as the state-of-the-art for prediction tasks (e.g., robotics, weather forecasting). However, deployment in safety-critical settings is hindered by a critical extrapolation hazard: driven by smoothness biases, flow models yield plausible outputs even for off-manifold conditions, resulting in silent failures indistinguishable from valid predictions. In this work, we introduce Diverging Flows, a novel approach that enables a single model to simultaneously perform conditional generation and native extrapolation detection by structurally enforcing inefficient transport for off-manifold inputs. We evaluate our method on synthetic manifolds, cross-domain style transfer, and weather temperature forecasting, demonstrating that it achieves effective detection of extrapolations without compromising predictive fidelity or inference latency. These results establish Diverging Flows as a robust solution for trustworthy flow models, paving the way for reliable deployment in domains such as medicine, robotics, and climate science.

Figures

Figures reproduced from arXiv: 2602.13061 by Constantinos Tsakonas, Jean-Baptiste Mouret, Serena Ivaldi.

Figure 1
Figure 1. Figure 1: Conceptual overview of Diverging Flows. (Top) Flow Matching minimizes transport cost everywhere, forcing off￾manifold inputs (red) to converge silently. (Bottom) Diverging Flows breaks this symmetry by enforcing transport inefficiency for invalid conditions. This ensures valid conditions (c) induce optimal flows, while off-manifold inputs (c˜) trigger detectable divergence. curved trajectories due to their… view at source ↗
Figure 2
Figure 2. Figure 2: Vector Dynamics on Synthetic Manifolds. Two experiments, Probabilistic Regression (left), with horizon of 10 steps, and Conditional Generation (right). Top: Standard FM forces smooth convergence for all inputs, causing silent hallucinations. Bottom: Diverging Flows learns a conservative field: on-manifold inputs (green) follow optimal paths, while off-manifold queries (orange) trigger divergent flows, crea… view at source ↗
Figure 3
Figure 3. Figure 3: High-Fidelity Forecasts with Diverging Flows. (Top) Visual comparison of 6-hour temperature forecasts. Diverging Flows accurately captures complex thermal fronts and fine-grained dynamics with MSE and SSIM of 0.0004 and 0.961 respectively, demonstrating that the regularization for hallucination detection does not compromise the model’s capacity for precise physical regression. (Bottom) Comparison of the be… view at source ↗
Figure 4
Figure 4. Figure 4: Generative Quality in Cross-Domain Transfer. Sam￾ples mapping structural inputs (MNIST) to street-view imagery (SVHN). Diverging Flows preserves the semantic identity of the digit. This demonstrates that regularization targets invalid seman￾tics without restricting diversity. 4.4. Efficacy of Negative Sampling Strategies To isolate the contribution of our adversarial negative min￾ing, we compare Diverging … view at source ↗
Figure 5
Figure 5. Figure 5: Detection Landscape: Conditional Generation. The standard model (a) fails to distinguish the manifold from the background. Diverging Flows (b) successfully identifies the ambient space as invalid (red), accepting only the spiral support. (a) Flow Matching (b) Diverging Flows (Ours) [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Detection Landscape: Probabilistic Regression. Even in the regression task, where the condition represents a dynamic state xt, Diverging Flows (b) maintains a precise decision boundary around the valid trajectory path. D. Extended Conformal Analysis In the main text, we report the False Positive Rate (FPR) at a fixed error rate α = 0.05 (Target Coverage 95%). Here, we expand on the relationship between Tar… view at source ↗
Figure 7
Figure 7. Figure 7: Extended Style Transfer Samples. Pairs of Conditioning Inputs (MNIST) and Generated Outputs (SVHN). The model successfully maps the sparse structural guidance of the grayscale digit to the rich, multi-modal distribution of street view imagery. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Extended Weather Temperature Forecasting Samples. Randomly selected 6-hour temperature forecasts generated by Diverging Flows on the ERA5 validation set. The model generates physically consistent heatmaps that align with the ground truth dynamics. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

27 extracted references · 17 linked inside Pith

  1. [5]

    Choi, H., Jang, E., and Alemi, A. A. Waic, but why? gen- erative ensembles for robust anomaly detection.arXiv preprint arXiv:1810.01392,

    Pith/arXiv arXiv

  2. [7]

    2023.3282993

    doi: 10.1109/ACCESS. 2023.3282993. Fawcett, T. An introduction to roc analysis.Pattern recog- nition letters, 27(8):861–874,

    arXiv 2023

  3. [9]

    L., Wang, F.-Y ., Herrera-Viedma, E., and Herrera, F

    Herrera-Poyatos, A., Del Ser, J., de Prado, M. L., Wang, F.-Y ., Herrera-Viedma, E., and Herrera, F. Responsible artificial intelligence systems: A roadmap to society’s trust through trustworthy ai, auditability, accountability, and governance.arXiv preprint arXiv:2503.04739,

    arXiv

  4. [10]

    J., H´olm, E., Janiskov´a, M., Keeley, S., Laloyaux, P., Lopez, P., Lupu, C., Radnoti, G., de Rosnay, P., Rozum, I., Vamborg, F., Villaume, S., and Th ´epaut, J.-N

    Hersbach, H., Bell, B., Berrisford, P., Hirahara, S., Hor´anyi, A., Mu˜noz-Sabater, J., Nicolas, J., Peubey, C., Radu, R., Schepers, D., Simmons, A., Soci, C., Abdalla, S., Abel- lan, X., Balsamo, G., Bechtold, P., Biavati, G., Bidlot, J., Bonavita, M., De Chiara, G., Dahlgren, P., Dee, D., Diamantakis, M., Dragani, R., Flemming, J., Forbes, R., Fuentes, ...

    1999

  5. [13]

    Kim, B., Kim, J., Kim, J., and Ye, J. C. Generalized consis- tency trajectory models for image manipulation.arXiv preprint arXiv:2403.12510,

    Pith/arXiv arXiv

  6. [14]

    Flow matching with gaussian process priors for probabilistic time series forecasting.arXiv preprint arXiv:2410.03024,

    Kollovieh, M., Lienen, M., L ¨udke, D., Schwinn, L., and G¨unnemann, S. Flow matching with gaussian process priors for probabilistic time series forecasting.arXiv preprint arXiv:2410.03024,

    Pith/arXiv arXiv

  7. [15]

    Enhancing the reliability of out-of-distribution image detection in neural networks

    Liang, S., Li, Y ., and Srikant, R. Enhancing the reliability of out-of-distribution image detection in neural networks. arXiv preprint arXiv:1706.02690,

    Pith/arXiv arXiv

  8. [19]

    W., Gorur, D., and Lakshminarayanan, B

    Nalisnick, E., Matsukawa, A., Teh, Y . W., Gorur, D., and Lakshminarayanan, B. Do deep generative models know what they don’t know?arXiv preprint arXiv:1810.09136,

    Pith/arXiv arXiv

  9. [20]

    Y ., et al

    Netzer, Y ., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A. Y ., et al. Reading digits in natural images with unsu- pervised feature learning. InNIPS Workshop on Deep Learning and Unsupervised Feature Learning, volume 2011, pp

    2011

  10. [21]

    Extremum flow matching for offline goal conditioned reinforcement learning.arXiv preprint arXiv:2505.19717,

    Rouxel, Q., Donoso, C., Chen, F., Ivaldi, S., and Mouret, J.- B. Extremum flow matching for offline goal conditioned reinforcement learning.arXiv preprint arXiv:2505.19717,

    Pith/arXiv arXiv

  11. [22]

    F., and Luque, J

    Serr`a, J., ´Alvarez, D., G´omez, V ., Slizovskaia, O., N´u˜nez, J. F., and Luque, J. Input complexity and out-of- distribution detection with likelihood-based generative models.arXiv preprint arXiv:1909.11480,

    Pith/arXiv arXiv 1909

  12. [23]

    Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., and Ganguli, S

    doi: 10.1109/IJCNN60899.2024.10650768. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., and Ganguli, S. Deep unsupervised learning using nonequi- librium thermodynamics. InInternational Conference on Machine Learning, pp. 2256–2265. PmLR,

    arXiv 2024

  13. [24]

    P., and Bovik, A

    Wang, Z., Simoncelli, E. P., and Bovik, A. C. Multiscale structural similarity for image quality assessment. InThe thrity-seventh asilomar conference on signals, systems & computers, 2003, volume 2, pp. 1398–1402. IEEE,

    2003

  14. [26]

    Unlike simple shapes, a spiral represents a stiff geometric structure where the optimal transport path (a straight line) often intersects regions that are off-manifold

    Note on the Spiral Benchmark.The 2D spiral is a canonical benchmark in the study of Neural ODEs and continuous-time dynamical systems (Chen et al., 2018). Unlike simple shapes, a spiral represents a stiff geometric structure where the optimal transport path (a straight line) often intersects regions that are off-manifold. In the context of probabilistic r...

    2018

  15. [64]

    To ensure stability, we use gradient clipping (norm 1.0) and an Exponential Moving Average (EMA) of model weights with decay0.9999. B.3. Cross-Domain Style Transfer Details Dataset.We perform a cross-domain translation task mapping grayscale MNIST digits ( 1×32×32 ) to RGB SVHN digits (3×32×32). Model Architecture.We use a U-Net backbone similar to the we...

    2024

  16. [2000]

    π0: A vision-language-action flow model for general robot control.arXiv preprint arXiv:2410.24164,

    Black, K., Brown, N., Driess, D., Esmail, A., Equi, M., Finn, C., Fusai, N., Groom, L., Hausman, K., Ichter, B., et al. π0: A vision-language-action flow model for general robot control.arXiv preprint arXiv:2410.24164,

    Pith/arXiv arXiv

  17. [2004]

    Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms.arXiv preprint arXiv:1708.07747,

    Xiao, H., Rasul, K., and V ollgraf, R. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms.arXiv preprint arXiv:1708.07747,

    Pith/arXiv arXiv

  18. [2006]

    Deep anomaly detection with outlier exposure.arXiv preprint arXiv:1812.04606,

    Hendrycks, D., Mazeika, M., and Dietterich, T. Deep anomaly detection with outlier exposure.arXiv preprint arXiv:1812.04606,

    Pith/arXiv arXiv

  19. [2010]

    Alphafold meets flow matching for generating protein ensembles

    Jing, B., Berger, B., and Jaakkola, T. Alphafold meets flow matching for generating protein ensembles. InNeurIPS 2023 Generative AI and Biology (GenBio) Workshop,

    2023

  20. [2017]

    T., Ben-Hamu, H., Nickel, M., and Le, M

    Lipman, Y ., Chen, R. T., Ben-Hamu, H., Nickel, M., and Le, M. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747,

    Pith/arXiv arXiv

  21. [2018]

    Deep learning for classical japanese literature.arXiv preprint arXiv:1812.01718,

    Clanuwat, T., Bober-Irizar, M., Kitamoto, A., Lamb, A., Yamamoto, K., and Ha, D. Deep learning for classical japanese literature.arXiv preprint arXiv:1812.01718,

    Pith/arXiv arXiv

  22. [2020]

    Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., and Hochreiter, S

    doi: https://doi.org/10.1002/qj.3803. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., and Hochreiter, S. Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advances in Neural Information Processing Systems, 30,

    doi:10.1002/qj.3803

  23. [2021]

    Towards deep learning models resistant to adversarial attacks.arXiv preprint arXiv:1706.06083,

    10 Diverging Flows: Detecting Extrapolations in Conditional Generation Madry, A., Makelov, A., Schmidt, L., Tsipras, D., and Vladu, A. Towards deep learning models resistant to adversarial attacks.arXiv preprint arXiv:1706.06083,

    Pith/arXiv arXiv

  24. [2022]

    T., Lopez-Paz, D., Ben-Hamu, H., and Gat, I

    Lipman, Y ., Havasi, M., Holderrieth, P., Shaul, N., Le, M., Karrer, B., Chen, R. T., Lopez-Paz, D., Ben-Hamu, H., and Gat, I. Flow matching guide and code.arXiv preprint arXiv:2412.06264,

    Pith/arXiv arXiv

  25. [2023]

    Chalapathy, R

    doi: 10.1177/20539517231169731. Chalapathy, R. and Chawla, S. Deep learning for anomaly detection: A survey.arXiv preprint arXiv:1901.03407,

    Pith/arXiv arXiv 1901

  26. [2024]

    J., Akhound-Sadegh, T., Huguet, G., Fatras, K., Rector-Brooks, J., Liu, C.-H., Nica, A

    Bose, A. J., Akhound-Sadegh, T., Huguet, G., Fatras, K., Rector-Brooks, J., Liu, C.-H., Nica, A. C., Korablyov, M., Bronstein, M., and Tong, A. Se (3)-stochastic flow matching for protein backbone generation.arXiv preprint arXiv:2310.02391,

    Pith/arXiv arXiv

  27. [2025]

    N., Barber, R

    Angelopoulos, A. N., Barber, R. F., and Bates, S. Theoreti- cal foundations of conformal prediction.arXiv preprint arXiv:2411.11824,

    Pith/arXiv arXiv