acceptodds
Under review as a conference paper at ICLR 2027

Match the Radius, Learn the Angles: Radial-Angular Flow Matching

Abstract

Flow Matching commonly relies on Gaussian sources and Euclidean conditional paths. For data with non-Gaussian amplitude statistics, this introduces a source-induced radial mismatch: the learned transport must reshape the distribution of norms while simultaneously modeling the directional structure of the data. We propose Radial-Angular Flow Matching (RAFM), which matches the target radial law at the source and preserves it along spherical conditional paths, separating radial estimation from directional transport. We further show that, although this construction removes radial motion from the path, conventional velocity regression still inherits the scale of the data. This motivates a scale-free angular formulation whose targets are uniformly bounded independently of the radial distribution, while still recovering the corresponding Flow Matching velocity field after radial rescaling. Our analysis characterizes the Gaussian radial mismatch removed by source matching and shows that, unlike velocity targets, angular targets require no finite radial moments. Across controlled heavy-tailed benchmarks and real vector, image, and audio data, ablations isolate the contributions of source matching, path geometry, and angular regression. The results show that these components provide complementary benefits while retaining the simulation-free training framework of Flow Matching.

Primary area:
generative models
Keywords:
Flow Matching, generative modeling, heavy-tailed distributions, non-Gaussian source distributions, radial-angular transport, spherical geometry, continuous normalizing flows

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_nHRhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

RAFM uses an empirical radial source and spherical paths to separate radial estimation from directional transport in flow matching. RAFM-Ang normalizes velocity targets by radius to obtain bounded angular targets, motivated by the radial-moment requirements of conventional velocity regression. The paper provides theoretical analyses and evaluates the method on heavy-tailed synthetic data, real vector datasets, and image and audio generation.

Critical strengths
  • Clear mathematical motivation. The radial-angular decomposition provides a coherent motivation for the method. The analysis makes the radial weighting in velocity regression explicit and establishes bounded angular targets, which is relevant to heavy-tailed settings.

  • Informative controlled comparisons. The matched-source Euclidean FM and fixed-spherical baselines provide useful comparisons for assessing the source and transport design. The Ang-R² ablation directly tests loss weighting while keeping the angular parameterization and sampling field fixed.

  • Clear presentation. The paper is well organized and easy to follow. The progression from radial source matching to spherical transport and angular regression makes the motivation and method clear.

Critical weaknesses
  • Modest methodological contribution. RAFM combines empirical radial matching with spherical flow matching and uses angular regression to avoid the implicit R² weighting of velocity regression. The spherical construction builds on RFM, while norm-preserving generation has a precedent in MSGM. The analysis is useful, but I find the methodological contribution incremental.

  • Unclear quality–cost trade-off. RAFM has substantially lower reported training cost than MSGM, but Table 4 shows that RAFM-Ang requires roughly 1.9–3.5× the training time of Gaussian FM on Student-t/PIV. The comparisons use the same number of updates. It is unclear whether the gains persist when Gaussian FM is given the same training-time budget.

  • Limited image-generation evaluation. Natural-image generation is evaluated only on ImageNette with one DC-AE/SiT configuration. It is unclear whether the gains over Gaussian FM extend to other image datasets or model configurations.

Questions for authors
  1. Does RAFM-Ang still outperform Gaussian FM on Student-t/PIV when both methods are given the same wall-clock training budget?

  2. Could the authors compare RAFM-Ang and Gaussian FM on another natural-image generation benchmark, such as ImageNet?

1 comment

We thank the reviewer for the constructive feedback. RAFM primarily targets distributions with challenging radial structure, especially heavy-tailed or strongly non-Gaussian data. Our additional experiments confirm its benefits at equal wall-clock time and demonstrate scalability to ImageNet-256.

W1 (contribution). We acknowledge the precedents of spherical paths (RFM) and norm-preserving generation (MSGM). Our contribution is scale-free angular regression: standard velocity regression implicitly weights the angular loss by R2R^2, whereas RAFM-Ang removes this weighting, has targets bounded by π\pi independently of radial moments, and recovers the FM velocity through rescaling. Training remains simulation-free, unlike MSGM.

A new heavy-tail sweep (Student-t, ν=1.5\nu=1.5–55, d=16/32d=16/32) confirms the mechanism: RAFM-Ang's loss max/median is ≤3.3, versus 34–330 for Gaussian FM, Matched FM, RAFM-Vel, and Student-t FM at ν=1.5\nu=1.5. Restoring the R2R^2 weighting (Ang-R2R^2) restores the spikes. RAFM-Ang beats Gaussian FM in all 8 settings and achieves the best Sliced W1W_1 in 5. Further comparisons with spherical controls and MSGM are discussed in our response to reviewer n2NE, with supporting results in its follow-up comment (Tables R1–R3).

W2/Q1 (equal wall-clock). On MI210 GPUs, Gaussian FM was allowed the same training time as RAFM-Ang's 10k updates, completing approximately 11.5k–13.8k updates in the first three settings. Gaussian FM checkpoints were evaluated every 500 updates. Sliced W1W_1 ↓, mean ± SD over 3 seeds:

SettingBudgetRAFM-AngGaussian FM lastGaussian FM best
Student-t d1635 s0.341 ± 0.0681.042 ± 0.2030.663 ± 0.216
Student-t d3233 s0.571 ± 0.0831.937 ± 0.5151.106 ± 0.079
PIV d6430 s0.0277 ± 0.00190.0464 ± 0.00410.0432 ± 0.0013
PIV d256, width 51227.1 s0.0263 ± 0.00500.0521 ± 0.00300.0498 ± 0.0047

RAFM-Ang outperforms even Gaussian FM's best checkpoint within the budget in all four settings. Student-t uses freshly drawn samples, explaining differences from the original manuscript results.

The 1.9–3.5× training-time ratios reported in Table 4 of the submitted paper were measured on an RTX 2000 Ada. On MI210, RAFM-Ang's per-update overhead is 15–38%.

W3/Q2 (ImageNet-256). We additionally evaluated class-conditional ImageNet-256 using DC-AE latents and SiT-B/1, training for 400k updates with 3 seeds per method. FID-10k ↓ (mean ± SD):

UpdatesGaussian FMRAFM-Ang
50k74.68 ± 0.3473.19 ± 0.46
100k57.61 ± 0.3157.01 ± 0.39
150k49.81 ± 0.3049.49 ± 0.35
200k45.15 ± 0.4144.70 ± 0.43
250k41.98 ± 0.2041.19 ± 0.21
300k39.69 ± 0.1838.92 ± 0.28
350k37.73 ± 0.1836.99 ± 0.15
400k36.37 ± 0.3135.34 ± 0.17

RAFM-Ang achieves a slightly lower mean FID at every checkpoint, reaching 35.34 versus 36.37 at 400k updates, with comparable training times (8.98 h versus 8.92 h). These results demonstrate scalability to high-dimensional image latents without substantial computational overhead. We emphasize that our primary contribution concerns challenging non-Gaussian radial distributions, where the gains are considerably more pronounced.

Revisions. We will incorporate these experiments, report equal-time budgets and standard deviations, clarify the PIV architectural limitation, and strengthen the discussion of RAFM's scope and relationship to prior work.

We hope these additional experiments and clarifications address the reviewer's concerns and would be happy to address any further questions.

Reviewer_NGDGOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper argues that a Gaussian source in Flow Matching induces a radial mismatch when data norms are non-Gaussian (e.g., heavy-tailed). It proposes Radial-Angular Flow Matching (RAFM), which draws the source radius from the empirical target radial law with a uniform direction and connects each pair by a radius-preserving spherical geodesic. Optionally (RAFM-Ang), it regresses the radius-normalized angular velocity, whose target is bounded by π independently of radial moments. The theory includes a polar KL decomposition, target-moment results, and guarantees for the empirical radial source. Experiments follow an ablation ladder (Gaussian FM → matched-source FM → RAFM-Vel → RAFM-Ang) on synthetic Student-t, PIV, finance, weather, ImageNette DC-AE latents, FSD50K, and AudioMNIST, with mechanism controls and t-Flow/MSGM baselines.

Critical strengths
  • The radial/angular decomposition is simple and well motivated. It adds negligible cost and keeps training simulation-free.
  • The bounded angular target is a clean idea. The Ang-R² control shows that the output parameterization matters beyond loss weighting.
  • The ablation ladder is careful, and negative results (Finance/Weather, SEVIR, post-hoc gain rescaling) are reported openly.
Critical weaknesses
  • Radii are resampled from training norms and preserved exactly, so RAFM cannot generate radii beyond the training maximum. This conflicts with the heavy-tail motivation, and no tail-specific evaluation is provided.
  • On ImageNette and FSD50K, matched-source FM and RAFM-Vel are worse than Gaussian FM; only RAFM-Ang helps. The claimed complementarity of the components is therefore weakly supported in high dimensions, and the ImageNette FIDs (150–175) are far from a practical regime.
  • The external baselines look like failed adaptations: t-Flow reaches Sliced W1 up to 196 and an energy score of 13,988 on SEVIR, and MSGM is at chance (0.100) on AudioMNIST. A simple radial-Gaussianization baseline is also missing.
Questions for authors
  1. How does RAFM compare with mapping radii monotonically to the χ_d law, x↦Fχd−1(F^R(∥x∥)) x/∥x∥x \mapsto F_{\chi_d}^{-1}(\hat F_R(\|x\|))\,x/\|x\|, training standard Gaussian FM, and inverting the map after sampling?
1 comment

We thank the reviewer for the constructive feedback and the radial-Gaussianization (RG) suggestion. We will incorporate these new results into the paper.

W1. Finite radial support and tail evaluation. Our empirical radial construction preserves the observed radius distribution by design. To assess its tail fidelity, we evaluated Student-t distributions (ν ∈ {1.5, 2, 3, 5}, d ∈ {16, 32}; 3 seeds; 10⁶ reference samples). At the 99.9th radial quantile, RAFM achieves 0.04–0.19 relative error, versus 0.24–0.97 for Gaussian FM. Its exceedance rate is 0.0008–0.0013 (ideal: 0.001), versus 0 for Gaussian FM. Matched-source FM yields comparable results (quantile error 0.02–0.29), confirming the contribution of radial matching. The reference probability beyond the training maximum is at most 5.4×10⁻⁵, and none of the evaluated methods generated samples beyond that maximum. We will discuss this finite-support property alongside the new tail results.

W2. Complementarity in high dimensions. The components address distinct aspects of transport and do not necessarily yield additive improvements. On FSD50K, RAFM-Ang achieves 74.79 FAD, versus 86.30 (Gaussian FM), 89.09 (Matched FM), and 90.95 (RAFM-Vel), highlighting the importance of angular regression.

We additionally evaluated ImageNet-256 (DC-AE latents, SiT-B/1, 400k updates, 3 seeds, no guidance). At 400k updates, RAFM-Ang achieves 35.34 ± 0.17 FID, versus 36.37 ± 0.31 for Gaussian FM. RAFM-Ang also performs better at the earlier evaluated checkpoints, from 50k to 350k updates. These results extend our evaluation beyond ImageNette and demonstrate competitive image generation at a substantially larger scale.

W3. External baselines. t-Flow: We reproduced its poor results (e.g., SEVIR: 13,991 vs 13,988) using the authors' formulation with documented adaptations. Its field (x−ε^)/t(x-\hat{\varepsilon})/t amplifies prediction errors by 1/t1/t, reaching 21–771× the true velocity near t=10−3t=10^{-3}. A near-Gaussian control fails similarly. We therefore added Student-t-source FM with velocity regression as a direct baseline; RAFM-Ang achieves 3 wins and 3 ties across six vector benchmarks (Table R3 in our follow-up comment to reviewer n2NE).

MSGM: We used the authors' public sparse implementation and reproduced the vector results. On AudioMNIST (d = 16,254), its generator mixes too slowly to reach the assumed prior (required integrated rate ≈6.7×10⁶, versus 5 in the schedule). This plausibly explains the chance-level accuracy but does not establish that our adaptation is optimal. We do not treat this failure as evidence of RAFM's superiority.

Q1. Radial Gaussianization. We implemented the proposed ECDF mapping to the Gaussian radial law, trained standard Gaussian FM, and inverted the map before evaluation. Results are mean ± SD over 3 seeds:

SettingGaussian FMRGRAFM-Ang
Student-t d160.682 ± 0.1560.284 ± 0.0190.292 ± 0.015
Student-t d321.750 ± 0.0300.745 ± 0.0870.498 ± 0.073
PIV d640.0459 ± 0.00410.0304 ± 0.00440.0275 ± 0.0025
PIV d2560.451 ± 0.004†0.132 ± 0.0010.0223 ± 0.0015
Finance0.193 ± 0.0240.226 ± 0.0120.174 ± 0.002
Weather0.094 ± 0.0100.105 ± 0.0100.078 ± 0.002
ImageNette, FID ↓162.2 ± 1.1178.1 ± 4.1150.0 ± 0.6
AudioMNIST, accuracy ↑0.732 ± 0.0100.851 ± 0.0100.780 ± 0.020
AudioMNIST, energy KS ↓0.114 ± 0.0110.521 ± 0.0130.022 ± 0.000

Vectors: Sliced W1 ↓; Student-t d32 uses batch 256 (paper: 4096); ImageNette uses 40k updates. †Gaussian FM's width-128 bottleneck (see our follow-up comment to reviewer n2NE). Bold indicates the best numerical mean.

RG beats Gaussian FM on four of six vector datasets but never significantly outperforms RAFM-Ang (4 wins, 2 ties for RAFM-Ang). Across eight heavy-tail configurations, RAFM-Ang achieves 5 wins and 3 ties against RG. On AudioMNIST, RG obtains higher classification accuracy, whereas RAFM better matches the energy distribution.

In high dimensions, the concentration of the χd\chi_d radial law makes the inverse RG mapping sensitive to radial errors. RAFM avoids this transformation by preserving data-matched radii throughout transport.

We thank the reviewer again for the constructive feedback, which helped us strengthen both the empirical evaluation and the discussion of RAFM's limitations. We hope these additional results address your concerns and would be happy to clarify any remaining questions.

Reviewer_n2NEOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The paper proposed RAFM, that matches the target radial law to the source and preserve it along the spherical conditional paths, and separates radial estimation from directional transport. Furthermore, the paper showed while the construction keeps the radius constant along the path, conventional velocity regression still inherits the scale of the data, motivating scale-free angular formulation. The empirical results show the proposed compnents provide complementary benefits while retaining the simulation-free training framework of Flow Matching.

Critical strengths
  1. The result on bounding target appears correct, and provides a perspective on radial weighting that is implicit in conventional flow matching, i.e. the velocity regression.
  2. The ablation study is comprehensive.
  3. The results support a benefit from spherical paths beyond source matching alone.
Critical weaknesses
  1. The writing of the paper seems convoluted and hard to follow.
  2. The methodological part is somewhat incremental. Works such as Multiplicative Diffusion Models already use a latent distribution combining data's radial law with independent uniform directions, whereas Riemannian Flow Matching already has simulation-free geodesic support on the spheres. The main part here seems to be combining the matched-radius coupling, and the scale-normalised regression objective. Positioning against other related works, such as Geometry-aware flow matching, needs to be done properly.
  3. It's also unclear what the practical advantage is of the proposed objective against the theoretical guarantees. For e.g I believe Theorem A.11 depends on the weighted angular loss, which can be minimised through training, but it's not clear if this would control the Euclidean generation error in the heavy tailed regimes that motivates the method.
  4. There seems to be severe failure of several baselines, could authors comment on that?
Questions for authors

See Weaknesses above

2 comments

Thank you for the review. Results and diagnostics from the additional experimental results are provided in a follow-up comment below (Tables R1–R3).

Novelty. The ingredients have precedents, but their combination addresses a different learning problem. MSGM (Gruhlke et al., 2026) uses a source with matched radii and uniform directions, but requires simulated diffusion trajectories. Riemannian FM (Chen & Lipman, 2024) provides simulation-free spherical paths. RAFM combines both. This changes the objective: since vt=RAtv_t=RA_t and ∥At∥≤π\|A_t\|\le\pi, ordinary velocity regression still weights angular errors by R2R^2 (Eq. 12). RAFM-Ang removes that weighting. The Ang-R2R^2 ablation brings back the loss spikes (new, Table R2: max/median 7–290, versus 1.1–3.3 for RAFM-Ang at ν≤2\nu\le2). RAFM is also substantially faster than MSGM, with lower Sliced W1W_1 on all six vector datasets (paper Table 1).

Geometry-Aware Image Flow Matching (Lee et al., ICML 2026). We will add this reference to Related Work. Their SFM operates on a single fixed-radius sphere, whereas RAFM preserves the full empirical radial distribution. With fixed RR, the R2R^2 weighting above is constant; with heavy-tailed variable radii, it is not. The radial share of transport cost is 1.0–1.3% on image latents versus 58–100% on Student-t and PIV. Our fixed-spherical control isolates this component and never beats RAFM-Ang (new, Table R3: 4 wins and 2 ties on vectors; ImageNette, paper Table 2: FID 166.1 vs 150.0). Independent radii tie RAFM-Ang on 5/6 vector datasets (Table R3), suggesting that radial matching explains much of the gain. Dependence can also matter: on our SO(2) control, Sliced W1W_1 is 0.027 for RAFM-Ang versus 0.321 with independent radii, while the methods tie when radius and direction are independent.

Theorem A.11 and heavy tails. A.11 controls an R2R^2-weighted error and assumes a finite second radial moment, so it does not cover infinite-variance targets. We will retain A.11 and add a bound for the unweighted objective we actually train. Let εθ=LAng(θ)−inf⁡aLAng(a)\varepsilon_\theta=\mathcal L_{\rm Ang}(\theta)-\inf_a\mathcal L_{\rm Ang}(a). Under the same type of Lipschitz regularity as A.11, applied to the projected field used by the ODE, and with the exact radial source, we obtain

ER[W22(pθ(U∣R),pdata(U∣R))]≤e2Lεθ.\mathbb E_R\left[W_2^2\bigl(p_\theta(U\mid R),p_{\rm data}(U\mid R)\bigr)\right]\le e^{2L}\varepsilon_\theta.

This controls generated directions without radial moments, using excess-risk decomposition, the marginal continuity equation, and Grönwall. With the empirical radial source, for any M>0M>0 we also obtain

W1(pθ,pdata)≤eLW1(μ^R,n,μR)+MeLεθ+2E[R1{R>M}].W_1(p_\theta,p_{\rm data})\le e^L W_1(\hat\mu_{R,n},\mu_R)+Me^L\sqrt{\varepsilon_\theta}+2\mathbb E[R\mathbf1_{\{R>M\}}].

This needs only E[R]<∞\mathbb E[R]<\infty and therefore covers Student-t with 1<ν≤21<\nu\le2. We will provide proofs, including a log-polar extension, in the appendix. For ν≤2\nu\le2, Euclidean W2W_2 is infinite between the infinite-variance target and any finite-variance model; that is why the angular and W1W_1 bounds are relevant here. In our new ν=1.5\nu=1.5–5 sweep (Table R1), RAFM-Ang has the best Sliced W1W_1 in all four settings with ν≤2\nu\le2 (3.78 vs 14.0 for Gaussian FM at d=32,ν=1.5d=32,\nu=1.5), while RAFM-Vel is competitive or better at ν=5\nu=5.

Baseline failures. We further investigated these results to rule out implementation errors and understand their causes. The diagnostics point to limitations in the settings tested:

  • t-Flow: in our MLP adaptation, the field (x−ϵ^)/t(x-\hat\epsilon)/t amplifies noise-prediction errors by 1/t1/t (21–771 times the true velocity at t=10−3t=10^{-3}). A near-Gaussian control (ν=104\nu=10^4) fails similarly, so the Student-t prior is not the cause. We added direct-velocity Student-t FM as a separate baseline; it never significantly outperforms RAFM-Ang (Table R3).
  • MSGM: the authors' sparse implementation reproduces the PIV scores reported in our paper. For flattened AudioMNIST (d=16,254d=16{,}254), the sparse generator does not mix over the given schedule (estimated integrated rate needed: 6.7×1066.7\times10^6; available: 5.025). This explains the limitation of this sparse variant, not of multiplicative diffusion generally.
  • PIV d=256d=256: we found that the original width-128 MLP made Gaussian FM look worse than it should. At width 512, RAFM-Ang still beats Gaussian FM (Sliced W1W_1: 0.024 vs 0.062 at equal updates), but ties Matched FM and RAFM-Vel at equal updates and loses to them at equal wall-clock time. We will report the corrected comparison.

Clarity. We have aimed to make the paper easy to follow, from the radial source construction to the learning objective and theoretical results. We will further clarify the presentation and would welcome pointers to specific passages that were difficult to follow.

AuthorsOpenReview ↗

Additional Experimental Results and Diagnostics

The following tables provide supporting results for our response above. All experiments were conducted for this rebuttal. Tables R1 and R3 report Sliced W1W_1 (↓), mean ± sample SD over 3 seeds. Matched FM: RAFM's radius-matched source with straight paths. RG: radial Gaussianization. Ang-R2R^2: RAFM-Ang with each loss term weighted by R2R^2, restoring the weighting of velocity regression.

Table R1. Heavy-tail sweep (Student-t; samples drawn for this sweep, so the ν = 3 cells differ from paper Table 1; 10610^6-sample reference; bold: best mean)

CellGauss. FMMatched FMRGStudent-t FMRAFM-VelRAFM-AngAng-R2R^2
d16, ν=1.57.25 ± 1.123.22 ± 0.505.93 ± 1.123.64 ± 0.162.58 ± 0.291.92 ± 0.243.16 ± 0.09
d16, ν=22.94 ± 0.541.53 ± 0.271.79 ± 0.241.32 ± 0.241.14 ± 0.370.720 ± 0.0581.36 ± 0.33
d16, ν=31.59 ± 0.270.702 ± 0.1260.368 ± 0.0420.477 ± 0.0500.399 ± 0.0340.342 ± 0.0660.448 ± 0.022
d16, ν=50.356 ± 0.0360.294 ± 0.0770.259 ± 0.0430.243 ± 0.0830.189 ± 0.0400.249 ± 0.0310.246 ± 0.066
d32, ν=1.514.0 ± 1.57.17 ± 0.458.75 ± 1.9610.5 ± 5.35.25 ± 0.843.78 ± 0.086.71 ± 1.16
d32, ν=23.44 ± 0.251.85 ± 0.201.26 ± 0.172.34 ± 0.231.46 ± 0.091.10 ± 0.141.43 ± 0.05
d32, ν=31.92 ± 0.330.774 ± 0.2510.770 ± 0.0871.01 ± 0.110.483 ± 0.0610.558 ± 0.0830.664 ± 0.088
d32, ν=51.18 ± 0.330.518 ± 0.0310.445 ± 0.0380.549 ± 0.0470.309 ± 0.0390.314 ± 0.0410.351 ± 0.032

Over the 48 comparisons, RAFM-Ang has 34 wins, 13 ties and 1 loss (to RAFM-Vel at d16, ν = 5). A tie means a mean gap within 2 standard errors.

Table R2. Training stability at ν ≤ 2 (max/median training loss over the second half of training, largest seed; no run had a non-finite loss)

Methodd16, ν=1.5d16, ν=2d32, ν=1.5d32, ν=2
RAFM-Ang1.11.13.31.1
Ang-R2R^295142907.0
Gauss. FM802518026
Matched FM110123008.9
RAFM-Vel100153308.0
Student-t FM341632010
RG1.11.11.11.1

Table R3. Spherical controls and Student-t FM on the paper's vector benchmarks (L/T: loses to/ties RAFM-Ang)

DatasetFixed radiusIndep. radiiStudent-t FMRAFM-Ang
Student-t d160.647 ± 0.036 L0.294 ± 0.048 T0.468 ± 0.032 L0.292 ± 0.015
Student-t d320.793 ± 0.072 L0.505 ± 0.098 T0.885 ± 0.154 L0.498 ± 0.073
PIV d640.0435 ± 0.0019 L0.0380 ± 0.0009 L0.0336 ± 0.0049 T0.0275 ± 0.0025
PIV d2560.0240 ± 0.0035 T0.0254 ± 0.0034 T0.0225 ± 0.0032 T0.0223 ± 0.0015
Finance0.193 ± 0.006 L0.180 ± 0.005 T0.172 ± 0.006 T0.174 ± 0.002
Weather0.0825 ± 0.0059 T0.0823 ± 0.0052 T0.0822 ± 0.0037 L0.0775 ± 0.0016

Fixed radius: one sphere at the mean training radius, as in Lee et al. (2026). Indep. radii: the same model with radii drawn independently from the training law. Student-t ν = 3 on the paper's samples. Released configuration: PIV d256 at width 128, Student-t d32 at batch 256 (paper Table 1's d32 RAFM-Ang row used batch 4096).

Comments on the paper

1 comment
Area_Chair_reQFOpenReview ↗

Engage in discussion

The authors and reviewers are requested to engage in the discussion phase.

Authors: please respond to the review comments.

Reviewers: please review the authors’ rebuttal and consider revising your assessment/score.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.