Machine Learning · arXiv:2611.13311 v1 · submitted 9 Nov 2026 · 21 references
One Trajectory, Two Requirements: Accuracy and Local Stability in Euler Sampling for Flow Matching Policies
The paper says
“Flow matching robot policies commonly use a fixed Euler step count selected from offline accuracy sweeps.”
more of the abstract
At inference time, however, each policy input (visual observation, robot state, and language instruction) together with a source-noise draw defines a different ODE trajectory and sequence of local Jacobians. We ask whether a step count that accurately follows the nominal noise-to-action trajectory can still be too coarse for strongly contracting directions around it. An exact two-coordinate system and a controlled augmentation of six trained flow models show that trajectory accuracy and local Euler stability can impose different step requirements even when the original learned transport is unchanged. We then study a frozen π0.5-LIBERO checkpoint without modifying its weights or action dynamics. Across 30 policy inputs from 10 LIBERO-10 tasks, the Jacobian exhibits strong contracting directions near the action endpoint, yet no sampled mode violates the Euler non-expansion condition. On five of those inputs, accuracy sweeps require 4–9 Euler steps at 5% terminal error and 87–112 at 1%. These five paired cases are therefore consistent with accuracy-limited rather than stability-limited sampling under this local criterion, while the controlled experiments show why endpoint accuracy alone cannot rule out a separate stability requirement.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 11 Nov 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. A clean, well-scoped paper that separates accuracy from stability in Euler sampling for flow matching policies. The controlled experiments are the real contribution; the π0.5 analysis is a careful but limited case study. This line is wrong
3 major objections · 5 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 21 references say.
The paper says, at §5.3–§5.4, Table 3, and abstract
The review did not quote this passage.
Pith says: Objection 1 of 3 · major
The 2% and 1% accuracy grids (N=27–39 for 2%, N=87–112 for 1%) place Euler nodes at t=m/N as small as about 1/39 and 1/112, which is strictly inside the interval (0, 1/32) that the Jacobian sampling does not cover (the smallest sampled time is 1/32).
more of objection 1
The inference in §5.4 that 'the additional steps required at tighter tolerances are driven by trajectory accuracy rather than by a sampled Euler stability violation' is therefore not supported for the 2% and 1% rows: a mode in (0, 1/32) with χ←(t)>N could push the stability requirement above the reported 87–112. The flat trend G(t)≈0.51 visible in Table 8 (t=1, 1/2, 1/32) is suggestive, but it is exactly the extrapolation the authors decline to make elsewhere (b121). Please either evaluate Jacobians at t=1/64, 1/128, and at the actual nodes of the N=87–112 grids, or explicitly restrict the 'accuracy-limited' interpretation at 2% and 1% to what the sampled window supports.
The paper says, at §3.1, Eq. (4), and Table 3
The review did not quote this passage.
Pith says: Objection 2 of 3 · major
The reported N_acc values are 'persistent' thresholds: the smallest N such that E_n≤ε for every tested n≥N. Because endpoint error is allowed to be non-monotone in N, a single outlying E_n can inflate the persistent threshold well above the first-hitting step count, and the paper's headline escalation (about a 12× step increase from 5% to 1% tolerance, roughly 4–9 to 87–112 steps) could be partly a persistence artifact rather than a stiff-region effect.
more of objection 2
Please plot E_N versus N over N=1..128 for the five paired inputs (or tabulate first-hitting versus persistent thresholds), so the escalation that motivates the tighter-tolerance interpretation is transparent.
The paper says, at §5.3 and Table 7
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.81):
“Therefore, finding no violation does not guarantee that the stability condition holds along the entire trajectory.”
Pith says: Objection 3 of 3 · major
For the negative stability finding on the LIBERO policy, the coverage rests on six largest-real-part and six largest-magnitude eigenpairs at each of 33 times, with only four full-Jacobian checks. The full-Jacobian checks validate the Arnoldi procedure at those four states, which mitigates the eigenvalue-coverage concern, but the time coverage (33 points, with the (0,1/32) gap) remains the binding limitation for a negative claim.
more of objection 3
A small number of direct finite-perturbation experiments on the five paired inputs would convert 'no sampled-mode violation' into evidence a practitioner can act on. The paper itself notes such a test would provide stronger evidence (§5.4), so this is a request to supply it rather than a claim that the current statement is wrong.
This review says the paper's mathematics is wrong. Pith did not check that against the paper's source.
5 minor objections
The paper says, at §3.2, Eq. (7)
Pith says: minor objection 1 of 5
The text first calls δ an infinitesimal perturbation (Eq. 7) and then, two sentences later, discusses 'finite perturbations' in the local-stability caveat; consider explicitly distinguishing the infinitesimal eigenvalue analysis from the finite-perturbation question to avoid a reader misreading the scope of Eq. (8).
The paper says, at Table 4
Pith says: minor objection 2 of 5
The caption of Table 4 mixes several column concepts (N_base, N̂stab_base, N_insert, N_pred=N_joint, N_full, R−z) without pointing the reader to the definitions in §4.2 and Appendix B; one sentence clarifying that N_full is the threshold for full endpoint error while N_joint includes non-expansion would help.
show 3 more
The paper says, at Figure 2 caption
Pith says: minor objection 3 of 5
Figure 2 uses forward time s while the rest of the paper uses native time t; the one-line reminder appears in the equation block but not in the figure caption, and a reader skimming the figures could misread the time axis.
The paper says, at §4.2, ε=0.01 column
Pith says: minor objection 4 of 5
The reported accuracy budgets at ε=0.01 (58.33±10.69 for Eight-Gaussians and 108.33±1.15 for Langevin) have a large seed-dependent spread for Eight-Gaussians; a brief comment on whether this spread reflects a particular checkpoint or a persistence-threshold jump would preempt a concern about outlier-driven averages.
The paper says, at Notation, Appendix E
Pith says: minor objection 5 of 5
Appendix E defines G(t) but the appendix table does not list R_z(N) from Experiment 1 or R^ref_N from Eq. (15); adding these would make the notation self-contained for a reader who starts at the experiments.
What would settle it
To further verify the robustness of the stability analysis, one could perform a direct perturbation experiment on a subset of the 30 policy inputs: introduce a small finite perturbation (e.g., 1e-4 relative norm) to the reference trajectory at the time of maximum G (t=1/32) and compare the evolution of the perturbation under the Euler sampler (N=10) versus a refined solver (e.g., RK4-1024). If the perturbation grows under Euler but contracts under the refined solver, it would indicate a stability limitation not captured by the eigenvalue analysis. If both contract, the eigenvalue-based conclusion is strengthened.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 11 Nov 2026
- Reader holds up, with conditions
- Skeptic agrees
- Editor wrote the verdict above
- Referee major revision: 3 major objections, 5 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
The paper does something useful and does it honestly. It takes a standard numerical-analysis distinction—accuracy vs. stability—and asks whether it matters for flow matching policies. The answer from the controlled experiments (Sections 4.1 and 4.2) is a clear yes: you can have a trajectory that is accurate at the endpoint while Euler amplifies a contracting mode. The two-coordinate system is exactly solved, the augmentation of trained flows is clean, and the results match the theory. This part is the real contribution.
The π0.5 analysis (Section 5) is a careful case study, not a general claim. The authors find no sampled stability violation across 30 inputs, and the accuracy budgets are larger than any stability concern. They are upfront about the limits: 30 inputs, one noise draw, five paired accuracy sweeps, local stability test, finite eigenvalue sampling. The paper does not overclaim. It says the policy is "consistent with accuracy-limited" sampling under this local criterion, which is exactly what the data support.
Soft spots: The stability analysis is local and linear. The authors acknowledge this, but it means the paper cannot rule out that non-normal transient growth or finite-perturbation effects matter in practice. The eigenvalue sampling (six largest real part, six largest magnitude) could miss modes that are not extreme in either sense but still cause trouble. The accuracy sweeps are only on five inputs, which is thin for a claim about the policy being accuracy-limited. The paper is honest about this, but it limits what a reader can conclude.
Who this is for: Researchers working on sampling efficiency for flow matching policies, especially those using Euler integration. The paper gives a clear framework for thinking about the two requirements and a method for checking stability. It does not claim to have found a practical problem in current policies, but it shows how to look for one.
Recommendation: Accept with minor revisions. The controlled experiments are solid. The π0.5 analysis is a reasonable demonstration of the method. The paper would be stronger with more accuracy sweeps and a discussion of non-normal effects, but it is already a fair and useful piece of work.
The strongest objection
The paper's central claim—that trajectory accuracy and local Euler stability can impose different step requirements in flow matching policies—is well-supported by the controlled experiments (Section 4). The analysis of the π0.5-LIBERO policy is appropriately cautious, with the authors explicitly acknowledging the limitations of the local stability test, the finite sampling of eigenvalues and times, and the small number of policy inputs and noise draws. The paper does not overclaim; it states that no stability violation was found among the sampled modes and times, and it does not assert that the policy is generally accuracy-limited. The weakest assumption identified by the reader—that the local stability analysis is sufficient—is already acknowledged by the authors as a limitation. The paper's own limitations section and multiple caveats throughout the text address this concern. Therefore, no additional load-bearing concern beyond what the authors themselves identify is apparent.
The reader's verdict is CONDITIONAL with high confidence, and the weakest assumption is that the local stability analysis is sufficient. The paper itself acknowledges this limitation explicitly in multiple places (e.g., 'This is a local stability test at a fixed point along the trajectory... it does not guarantee that finite perturbations will contract in coupled, time-varying dynamics', 'finding no violation does not guarantee that the stability condition holds along the entire trajectory', 'Limitations: ... we do not observe this limitation in the pretrained π0.5-LIBERO policy'). The controlled experiments are clean and support the central claim. The policy analysis is limited but the authors are transparent about the scope. The reader's concern is valid but already addressed by the paper's own caveats. Therefore, no change to the verdict is needed. The agreement is 'agree' because the reader correctly identifies the weakest assumption, and the paper's own limitations align with this.
The referee's report
The core conceptual contribution is solid and well executed: the paper cleanly separates trajectory accuracy from contracting-mode stability in Euler sampling, proves the separation on an exactly solvable system (Theorem 1), and demonstrates it on trained models via a controlled coordinate that provably leaves the learned transport unchanged (Proposition 1). The grid-aware sufficient condition G(t)≤1 (Proposition 2) is correct and is a genuinely useful framing, since an endpoint-inclusive Jacobian maximum would mis-state the sampler budget. The theory is shipped with closed-form proofs, and the controlled experiments are reproducible in structure (three seeds per target, 72 seed–tolerance–κ combinations as a consistency check). The LIBERO case study is exploratory and honestly scoped: RK4 refinement checks, full-Jacobian validation of the Arnoldi estimates (relative differences under 0.24%), and bitwise-identical reproducibility against the released checkpoint are all good empirical practice. If the empirical interpretation stands, the paper also offers an actionable message to practitioners choosing step budgets.
Pith's reading of the paper (a machine's summary, not the authors')
The paper argues that when a flow-matching robot policy is sampled with a fixed number of Euler steps, two distinct numerical requirements are at play: how accurately the nominal noise-to-action trajectory is followed (accuracy), and whether the step size is small enough to avoid amplifying locally contracting directions around that trajectory (stability). Using an exactly solvable two-coordinate system and six trained flow models with an artificially inserted contracting coordinate, the authors show that accuracy and stability can demand different step counts, and that endpoint accuracy alone cannot rule out a separate stability requirement. When they then examine a frozen π0.5-LIBERO checkpoint across 30 policy inputs, they find strong contracting directions near the action endpoint but no sampled contracting mode that violates the Euler non-expansion condition under their local criterion. On five inputs where both accuracy and stability were measured, the step budget was consistently set by accuracy, not stability. The practical lesson is that a fixed Euler budget should be interpreted through both the trajectory it follows and the local dynamics encountered along it.
What carries the argument. The key machinery is the local reverse-time Euler stability scale χ←(t,x;c)=max_{μ∈spec(J_v), Re(μ)>0} |μ|^2/(2 Re(μ)), which compares the contraction rate of each eigenmode of the state Jacobian J_v with the Euler step size h=1/N. A mode is non-expansive under Euler when N ≥ |μ|^2/(2 Re(μ)). The paper also introduces the scaled stability measure G(t)=t χ←(t), which provides a sufficient condition for non-expansion on the Euler grid when G(t)≤1 at all t∈(0,1].
Core claim. The paper establishes that trajectory accuracy and local Euler stability can impose different step requirements in flow matching policies, even when the learned transport is unchanged. In controlled settings, a contracting coordinate can force a larger Euler budget than accuracy alone would require, yet endpoint error can remain small just below the stability boundary. When applied to a pretrained π0.5-LIBERO policy, the paper finds that although the Jacobian exhibits strong contraction near the action endpoint, no sampled contracting mode violates the Euler non-expansion condition, and on the five inputs with paired accuracy sweeps the step budget is set by accuracy rather than stability.
The premise it rests on. The local stability analysis based on frozen-time Jacobian eigenvalues (the G(t) measure) is sufficient to determine whether stability imposes a separate step requirement, but the paper explicitly acknowledges that this test is local, does not capture transient growth from non-normal dynamics, and does not guarantee that finite perturbations contract in coupled, time-varying dynamics.
If the paper is right (Pith's inferences, not the authors')
For a given flow matching policy and tolerance, the required Euler budget is the maximum of the accuracy requirement and the stability requirement; ignoring either can lead to under-resolved sampling.
Endpoint accuracy sweeps alone are insufficient to guarantee that the Euler sampler is stable; a stability check based on local Jacobian eigenvalues is needed, especially near the action endpoint where contraction is strongest.
The stability requirement depends on the chosen tolerance: a system that is stability-limited at a loose tolerance can become accuracy-limited at a stricter tolerance, as shown in the controlled experiments.
The local stability criterion developed here can be applied to other flow matching policies beyond π0.5-LIBERO to determine whether their step budgets are accuracy- or stability-limited.
The finding that no sampled mode violates the Euler condition in π0.5-LIBERO suggests that the default 10-step Euler budget may be conservative with respect to stability, but this conclusion is limited to the 30 inputs and single noise draw tested.
The paper's controlled experiments show that stability can impose a separate requirement, but the pretrained policy analysis did not find a stability violation. This suggests that either (a) the π0.5-LIBERO policy has been designed or trained to avoid stiff contracting dynamics, or (b) the local eigenvalue test misses non-normal transient growth that could still cause instability. The paper acknow
A natural extension would be to perform direct perturbation experiments: inject a finite perturbation along a contracting eigenmode at a point on the trajectory and compare its evolution under Euler with a refined solver. The paper suggests this as future work.
The stability measure G(t) depends on the full Jacobian spectrum. The paper only samples six eigenvalues with largest real part and six with largest magnitude at 33 time points. A more exhaustive spectral analysis, especially near t=0, could reveal modes that were missed.
The paper's framework could be used to design adaptive step-size controllers that respond to both accuracy and stability requirements, rather than using a fixed budget chosen from offline sweeps.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Flow matching robot policies commonly use a fixed Euler step count selected from offline accuracy sweeps. At inference time, however, each policy input (visual observation, robot state, and language instruction) together with a source-noise draw defines a different ODE trajectory and sequence of local Jacobians. We ask whether a step count that accurately follows the nominal noise-to-action trajectory can still be too coarse for strongly contracting directions around it. An exact two-coordinate system and a controlled augmentation of six trained flow models show that trajectory accuracy and local Euler stability can impose different step requirements even when the original learned transport is unchanged. We then study a frozen π0.5-LIBERO checkpoint without modifying its weights or action dynamics. Across 30 policy inputs from 10 LIBERO-10 tasks, the Jacobian exhibits strong contracting directions near the action endpoint, yet no sampled mode violates the Euler non-expansion condition. On five of those inputs, accuracy sweeps require 4–9 Euler steps at 5% terminal error and 87–112 at 1%. These five paired cases are therefore consistent with accuracy-limited rather than stability-limited sampling under this local criterion, while the controlled experiments show why endpoint accuracy alone cannot rule out a separate stability requirement.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-11-11
The authors' main claim, as the classifier picked it out 0.76
“An exact two-coordinate system and a controlled augmentation of six trained flow models show that trajectory accuracy and local Euler stability can impose different step requirements even when the original learned transport is unchanged.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.72 · human participants 0.04 · controlled comparison 0.64 · randomized assignment 0.02 · states quantitative result 0.98 · compares with prior results 0.41 · mentions code or data availability 0.03
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Flow matching robot policies commonly use a fixed Euler step count selected from offline accuracy sweeps. | background 1.00 | 0.06 | 0.02 | asserted 1.00 |
| At inference time, however, each policy input (visual observation, robot state, and language instruction) together with a source-noise draw defines a different ODE trajectory and sequence of local Jacobians. | background 0.70 | 0.16 | 0.03 | asserted 1.00 |
| We ask whether a step count that accurately follows the nominal noise-to-action trajectory can still be too coarse for strongly contracting directions around it. | gap 0.94 | 0.12 | 0.08 | hedged possibility 0.69 |
| An exact two-coordinate system and a controlled augmentation of six trained flow models show that trajectory accuracy and local Euler stability can impose different step requirements even when the original learned transport is unchanged. | result 0.51 | 0.95 | 0.57 | asserted 1.00 |
| We then study a frozen $\pi_{0.5}$-LIBERO checkpoint without modifying its weights or action dynamics. | method 0.88 | 0.90 | 0.03 | asserted 0.90 |
| Across 30 policy inputs from 10 LIBERO-10 tasks, the Jacobian exhibits strong contracting directions near the action endpoint, yet no sampled mode violates the Euler non-expansion condition. | result 0.98 | 0.94 | 0.12 | asserted 1.00 |
| On five of those inputs, accuracy sweeps require 4--9 Euler steps at 5% terminal error and 87--112 at 1%. | result 1.00 | 0.93 | 0.08 | asserted 1.00 |
| These five paired cases are therefore consistent with accuracy-limited rather than stability-limited sampling under this local criterion, while the controlled experiments show why endpoint accuracy alone cannot rule out a separate stability requirement. | interpretation 1.00 | 0.91 | 0.53 | hedged possibility 0.95 |
What kind of paper the classifier reports this is: empirical study 0.87 · evidence from observation or experiment 0.54
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-11: 1 request, 6,316 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (8)
Figures are the authors', reproduced from the paper.
References (21)
- [1] K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, et al. π0: A vision-language- action flow model for general robot control.arXiv preprint arXiv:2410.24164, 2024. URL https://arxiv.org/abs/2410.24164 on PithPith's reviewarXiv
- [2] Physical Intelligence, K. Black, N. Brown, J. Darpinian, K. Dhabalia, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, et al. \pi_ {0.5}: a vision-language-action model with open-world 9 generalization.arXiv preprint arXiv:2504.16054, 2025 on PithPith's reviewarXiv
- [3] Q. Zhang, Z. Liu, H. Fan, G. Liu, B. Zeng, and S. Liu. FlowPolicy: Enabling fast and robust 3D flow-based policy via consistency flow matching for robot manipulation. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 14754–14762, 2025. doi: 10.1... on Pithdoi:10.1609/aaai.v39i14.33617
- [4] Y . Dai, H. Gu, T. Wang, Q. Cheng, Y . Zheng, Z. Qiu, L. Gong, W. Lou, and X. Zhou. Ac- tionflow: A pipelined action acceleration for vision language models on edge.arXiv preprint arXiv:2512.20276, 2025 on PithPith's reviewarXiv
- [5] C. Chi, S. Feng, Y . Du, Z. Xu, E. Cousineau, B. C. M. Burchfiel, and S. Song. Diffusion policy: Visuomotor policy learning via action diffusion. InRobotics: Science and Systems, on Pith
- [7] X. Hu, B. Liu, X. Liu, and Q. Liu. AdaFlow: Imitation learning with variance- adaptive flow-based policies. InAdvances in Neural Information Processing Systems, vol- ume 37, 2024. URL https://proceedings.neurips.cc/paper_files/paper/2024/ hash/fa9cfdb49f7de3dee54007f84c0745b1-... on Pith
- [8] Z. Fang, J. Wang, Y . Zhou, and Q. Shi. ProbeFlow: Training-free adaptive flow matching for vision-language-action models.arXiv preprint arXiv:2603.17850, 2026. URL https: //arxiv.org/abs/2603.17850 on PithPith's reviewarXiv
- [9] Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le. Flow matching for generative modeling. InInternational Conference on Learning Representations, 2023 on Pith
- [10] R. T. Q. Chen, Y . Rubanova, J. Bettencourt, and D. Duvenaud. Neural ordinary differential equations. InAdvances in Neural Information Processing Systems, volume 31, 2018 on Pith
- [11] A. Gupta, S. H. Lim, A. Yu, and N. B. Erichson. Sharpen your flow: Sharpness-aware sampling for flow matching.arXiv preprint arXiv:2605.11547, 2026. URL https://arxiv.org/abs/ 2605.11547 on PithPith's reviewarXiv
- [12] Z. Tan, W. Wang, A. L. Bertozzi, and E. K. Ryu. STORK: Faster diffusion and flow matching sampling by resolving both stiffness and structure-dependence. InInternational Conference on Learning Representations, 2026 on Pith
- [13] C. Tao and S.-K. Choi. On the role of strain and vorticity in numerical integration error for flow matching.arXiv preprint arXiv:2605.06680, 2026. URL https://arxiv.org/abs/2605. 06680 on PithPith's reviewarXiv
- [14] H. Xiao. From Euler to Dormand–Prince: ODE solvers for flow matching generative models. arXiv preprint arXiv:2605.00836, 2026. URLhttps://arxiv.org/abs/2605.00836 on PithPith's reviewarXiv
- [15] E. Hairer and G. Wanner.Solving Ordinary Differential Equations II: Stiff and Differential- Algebraic Problems. Springer, 1996 on Pith
- [16] L. N. Trefethen and M. Embree.Spectra and Pseudospectra: The Behavior of Nonnormal Matrices and Operators. Princeton University Press, 2005 on Pith
- [17] Physical Intelligence. OpenPI: Open-source models for robotics. Software repository, 2026. URL https://github.com/Physical-Intelligence/openpi. Implementation inspected at commit 215abfb217dbac7d5f1273282331b9b1866c0479. 10 on Pith
- [18] B. Liu, Y . Zhu, C. Gao, Y . Feng, Q. Liu, Y . Zhu, and P. Stone. LIBERO: Benchmarking knowledge transfer for lifelong robot learning. InAdvances in Neural Information Processing Systems, 2023. URLhttps://arxiv.org/abs/2306.03310 on PithPith's reviewarXiv
- [19] A. Tong, K. Fatras, N. Malkin, G. Huguet, Y . Zhang, J. Rector-Brooks, G. Wolf, and Y . Bengio. Improving and generalizing flow-based generative models with minibatch optimal transport. Transactions on Machine Learning Research, 2024 on Pith
- [20] N. Kornilov, P. Mokrov, A. Gasnikov, and A. Korotin. Optimal flow matching: Learning straight trajectories in just one step. InAdvances in Neural Information Processing Systems, 2024. 11 A Proofs and Details A.1 Proof of Theorem 1 Leth= 1/Nands k =kh. The exact increment of th... on Pith
- [2023] doi:10.15607/RSS.2023.XIX.026 on Pithdoi:10.15607/rss.2023.xix.026
show 1 more
- [2026] URLhttps://arxiv.org/abs/2609.21216 on PithPith's reviewarXiv
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment